CERIAS Tech Report 2004-51 ADAPTIVE AND HETEROGENEOUS MOBILE WIRELESS NETWORKS by Yi Lu Center for Education and Research in Information Assurance and Security, Purdue University, West Lafayette, IN 47907-2086
CERIAS Tech Report 2004-51
ADAPTIVE AND HETEROGENEOUS MOBILE WIRELESSNETWORKS
by Yi Lu
Center for Education and Research in Information Assurance and Security,
Purdue University, West Lafayette, IN 47907-2086
ADAPTIVE AND HETEROGENEOUS MOBILE WIRELESS NETWORKS
A Thesis
Submitted to the Faculty
of
Purdue University
by
Yi Lu
In Partial Fulfillment of the
Requirements for the Degree
of
Doctor of Philosophy
August 2004
ii
To my parents Zongfu Lu and Chunfang Cui and my wife Yuhui Zhong.
iii
ACKNOWLEDGMENTS
I would like to acknowledge the effort of my major advisor, Professor Bharat Bhar-
gava, in motivating and guiding me during the years that I spent in his research lab at
Purdue. Professor Bhargava’s expertise and insight are the keys for the success of his stu-
dents. I would like to thank Professor Michael Zoltowski for giving me ideas for wireless
networking research. I thank Professor Sonia Fahmy and Professor Aditya Mathur for
answering my questions and recommending research papers. I enjoyed the discussion of
research problems with Professor Dongyan Xu. Without the support and encouragement
of my advising committee, this thesis would not have been complete.
I thank my colleagues in the RAID lab for their help and support. Dr. Leszek Lilien,
Dr. Xiaoxin Wu, and Dr. Ahsan Habib gave me valuable comments on research ideas and
papers. The discussions with Weichao Wang and Mohamed Hefeeda have inspired me a
lot.
Finally, I greatly appreciate the help, support, and encouragement of my wife Yuhui
Zhong.
iv
TABLE OF CONTENTS
Page
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Problem statement. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Adaptive routing in mobile ad hoc networks. . . . . . . . . . . . 2
1.1.2 Large scale heterogeneous wireless networks. . . . . . . . . . . 3
1.2 Thesis contributions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Mobile ad hoc networks. . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Wireless networks with movable base stations. . . . . . . . . . . 5
1.3 Thesis organization. . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1 Destination-sequenced distance vector routing protocol (DSDV). . . . . 9
2.2 Ad hoc on-demand distance vector routing protocol (AODV). . . . . . . 10
2.3 Simulation environment. . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Mobility model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Study of ad hoc routing protocols. . . . . . . . . . . . . . . . . . . . . . . . 13
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.1 Problem statement. . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.2 Our contributions . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Related work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Correlation between topology change and mobility. . . . . . . . . . . . 15
3.4 Simulation settings and performance metrics. . . . . . . . . . . . . . . 17
3.5 Results and analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
v
Page
3.5.1 Varying maximum speed. . . . . . . . . . . . . . . . . . . . . 19
3.5.2 Varying number of connections. . . . . . . . . . . . . . . . . . 22
3.5.3 Dropped packets. . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.5.4 Varying number of mobile hosts. . . . . . . . . . . . . . . . . . 25
3.6 Further discussion about DSDV. . . . . . . . . . . . . . . . . . . . . . 26
3.6.1 Reduce broadcast interval of DSDV. . . . . . . . . . . . . . . . 26
3.6.2 Increase the queue length of DSDV. . . . . . . . . . . . . . . . 28
3.7 Congestion-aware routing protocol – CADV. . . . . . . . . . . . . . . . 28
3.7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.7.2 Preliminary results. . . . . . . . . . . . . . . . . . . . . . . . . 30
3.8 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4 Packet loss in ad hoc networks. . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Related work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 Simulation settings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3.1 Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3.2 Differentiated packet losses. . . . . . . . . . . . . . . . . . . . 36
4.4 Experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.4.1 Varying mobility and the number of connections. . . . . . . . . 37
4.4.2 Varying traffic load and traffic type. . . . . . . . . . . . . . . . 40
4.5 Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.6 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5 SAGA: self-adjusting congestion avoidance routing protocol. . . . . . . . . . 47
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Contention-based media access and congestion avoidance. . . . . . . . . 49
5.2.1 Characteristics of contention-based access to wireless channels. . 49
5.2.2 Ad hoc routing based on intermediate delay. . . . . . . . . . . . 50
5.3 Delay estimation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
vi
Page
5.3.1 The model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.3.2 Node with recent traffic. . . . . . . . . . . . . . . . . . . . . . 55
5.3.3 Node without recent traffic. . . . . . . . . . . . . . . . . . . . 55
5.3.4 Accuracy of delay estimation. . . . . . . . . . . . . . . . . . . 60
5.4 Self-adjusting congestion avoidance routing protocol. . . . . . . . . . . 62
5.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4.2 Operations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.5 Experimental evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.5.1 Performance metrics. . . . . . . . . . . . . . . . . . . . . . . . 71
5.5.2 Simulation and input parameters. . . . . . . . . . . . . . . . . . 71
5.5.3 Measurements and observations. . . . . . . . . . . . . . . . . . 73
5.5.4 Analysis and discussion. . . . . . . . . . . . . . . . . . . . . . 80
5.6 Related work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.7 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6 Hierarchical architecture for supporting movable base stations in wireless networks 88
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.2 Design considerations. . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.2.1 Asymmetric capacity and asymmetric responsibility. . . . . . . 89
6.2.2 Coordinated movement. . . . . . . . . . . . . . . . . . . . . . 90
6.2.3 Localized traffic. . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.2.4 Heterogeneous wireless networks. . . . . . . . . . . . . . . . . 90
6.3 Network architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.3.1 Definitions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.3.2 An example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.3.3 Basic operations. . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.4 Membership management. . . . . . . . . . . . . . . . . . . . . . . . . 98
6.4.1 Data structure. . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.4.2 Registration. . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
vii
Page
6.4.3 Leaving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.4.4 Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.5 Segmented membership-based group routing. . . . . . . . . . . . . . . 101
6.5.1 Data structure. . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.5.2 Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.6 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.7 Related work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.8 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7 Securing wireless networks with movable base stations. . . . . . . . . . . . . 109
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.1.1 Wireless network with movable base stations. . . . . . . . . . . 109
7.1.2 Security issues in WNMBS. . . . . . . . . . . . . . . . . . . . 110
7.2 Security objective and assumptions. . . . . . . . . . . . . . . . . . . . 112
7.3 Protection of network infrastructure. . . . . . . . . . . . . . . . . . . . 112
7.4 Authentication and key exchange. . . . . . . . . . . . . . . . . . . . . 115
7.4.1 Notations and protocol. . . . . . . . . . . . . . . . . . . . . . . 115
7.4.2 Correctness. . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.4.3 Security discussion. . . . . . . . . . . . . . . . . . . . . . . . 118
7.5 Secure roaming support. . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.5.1 Secure roaming support algorithm. . . . . . . . . . . . . . . . . 119
7.5.2 Mutual authentication between a mobile host and a FGA. . . . . 120
7.5.3 Fault-tolerant authentication. . . . . . . . . . . . . . . . . . . . 120
7.6 Computation overhead. . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.6.1 Overhead of secure packet forwarding. . . . . . . . . . . . . . . 121
7.6.2 Overhead of secure roaming support. . . . . . . . . . . . . . . . 123
7.7 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8 Conclusions and future work. . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
viii
Page
8.1.1 Study of ad hoc routing protocols. . . . . . . . . . . . . . . . . 127
8.1.2 Study of packet loss in ad hoc networks. . . . . . . . . . . . . . 128
8.1.3 Congestion avoidance routing protocol for ad hoc networks. . . . 129
8.1.4 Wireless networks with movable base stations. . . . . . . . . . . 131
8.1.5 Securing wireless networks with movable base stations. . . . . . 132
8.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.2.1 Congestion control in ad hoc networks. . . . . . . . . . . . . . 133
8.2.2 Trusted communication. . . . . . . . . . . . . . . . . . . . . . 134
8.2.3 Privacy-preserved communication. . . . . . . . . . . . . . . . . 135
LIST OF REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
VITA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
ix
LIST OF TABLES
Table Page
3.1 Power requirements. . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.1 Packet loss at MAC and network layers. . . . . . . . . . . . . . . . . . 36
5.1 Major constants of SAGA protocol. . . . . . . . . . . . . . . . . . . . 63
5.2 Simulation and input parameters. . . . . . . . . . . . . . . . . . . . . . 73
7.1 Encryption/decryption speed of block ciphers. . . . . . . . . . . . . . . 122
7.2 Speed of RSA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
x
LIST OF FIGURES
Figure Page
1.1 Network environments. . . . . . . . . . . . . . . . . . . . . . . . . . . 1
3.1 Topology change vs. mobility. . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Varying maximum speed. . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Varying number of connections. . . . . . . . . . . . . . . . . . . . . . 22
3.4 Dropped packets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.5 Varying number of mobile hosts. . . . . . . . . . . . . . . . . . . . . . 26
3.6 Performance comparison of different DSDV implementations. . . . . . . 27
3.7 Comparison of three protocols. . . . . . . . . . . . . . . . . . . . . . . 30
4.1 Packet loss for 4 packets/s CBR connections. . . . . . . . . . . . . . . . 38
4.2 Packet loss for 8 packets/s CBR connections. . . . . . . . . . . . . . . . 42
4.3 Packet loss for TCP connections. . . . . . . . . . . . . . . . . . . . . . 43
4.4 Shortest path and congestion. . . . . . . . . . . . . . . . . . . . . . . . 45
5.1 Network topology and flows. . . . . . . . . . . . . . . . . . . . . . . . 50
5.2 Select a route with presence of other connections. . . . . . . . . . . . . 51
5.3 Adapt to changes in traffic. . . . . . . . . . . . . . . . . . . . . . . . . 52
5.4 Adapt to changes in network topology. . . . . . . . . . . . . . . . . . . 52
5.5 Transmission of a unicast packet using RTS/CTS in the IEEE 802.11 stan-dard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.6 State transition of transmission procedure. . . . . . . . . . . . . . . . . 57
5.7 Comparison of estimated delay and measured delay. . . . . . . . . . . . 61
5.8 Data structure of the routing entry. . . . . . . . . . . . . . . . . . . . . 63
5.9 Delay estimate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.10 Algorithm for making an advertisement packet. . . . . . . . . . . . . . 66
5.11 Algorithm for route maintenance. . . . . . . . . . . . . . . . . . . . . 68
xi
Figure Page
5.12 Algorithm for handling broken links. . . . . . . . . . . . . . . . . . . . 69
5.13 POO traffic. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.14 10 CBR connections, low mobility. . . . . . . . . . . . . . . . . . . . . 74
5.15 10 CBR connections, high mobility. . . . . . . . . . . . . . . . . . . . 75
5.16 30 CBR connections, low mobility. . . . . . . . . . . . . . . . . . . . . 77
5.17 30 CBR connections, high mobility. . . . . . . . . . . . . . . . . . . . 78
5.18 POO traffic, low mobility . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.19 POO traffic, high mobility. . . . . . . . . . . . . . . . . . . . . . . . . 80
5.20 TCP traffic. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.1 Hierarchical mobile wireless network. . . . . . . . . . . . . . . . . . . 93
6.2 Hierarchy of groups. . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.3 After migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.4 Membership modification. . . . . . . . . . . . . . . . . . . . . . . . . 101
6.5 Membership table and routing table. . . . . . . . . . . . . . . . . . . . 102
6.6 SMGR routing algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.7 Protocol overhead versus number of mobile hosts. . . . . . . . . . . . . 105
7.1 Secure packet forwarding algorithm. . . . . . . . . . . . . . . . . . . . 114
7.2 Authentication and key exchange protocol. . . . . . . . . . . . . . . . . 116
7.3 Secure roaming support algorithm. . . . . . . . . . . . . . . . . . . . . 119
7.4 Mutual authentication protocol. . . . . . . . . . . . . . . . . . . . . . . 119
7.5 Topology of a WNMBS . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.6 Frequency of roaming requests. . . . . . . . . . . . . . . . . . . . . . . 125
xii
ABSTRACT
Lu, Yi. Ph.D., Purdue University, August, 2004. Adaptive and Heterogeneous MobileWireless Networks. Major Professor: Bharat Bhargava.
This dissertation investigates two research problems: (a) designing ad hoc routing pro-
tocols that monitor network conditions, select routes to satisfy routing requirements, and
adapt to network topology, traffic load, and congestion; (b) building an integrated infras-
tructure for heterogeneous wireless networks with movable base stations and developing
techniques for network management, routing, and security.
The experimental study of ad hoc routing protocols shows that the on-demand ap-
proach outperforms the proactive approach in less stressful situations, while the later one
is more scalable with respect to the network size. Mobility and congestion are the pri-
mary reasons for the packet loss for the on-demand and proactive approaches respectively.
Self-adjusting congestion avoidance (SAGA) routing protocol integrates the channel spa-
tial reuse with the multi-hop routing to reduce congestion. Using the intermediate delay
as the routing metric enables SAGA to bypass hot spots where contention is intense. An
estimate of the transmission delay is derived based on local information available at a
host. Comparison of SAGA with AODV, DSR, and DSDV shows that SAGA introduces
the lowest end-to-end delay. It outperforms DSDV in the measured metrics. SAGA can
sustain heavier traffic load and offers higher peak throughput than AODV and DSR. It
is shown that considerations of congestion and the intermediate delay can enhance the
routing performance significantly.
Hierarchical mobile wireless network is proposed to support wireless networks with
movable base stations. Mobile hosts are organized into hierarchical groups. An efficient
group membership management protocol is designed to support mobile hosts roaming
among different groups. Segmented membership-based group routing protocol takes ad-
xiii
vantage of the hierarchical structure and membership information to reduce overhead. A
secure packet forwarding algorithm is designed to protect the network infrastructure. The
roaming support algorithm cooperates with the proposed mutual authentication protocol to
secure both the foreign group and the mobile host. The evaluation shows that the compu-
tation overhead of the secure packet forwarding is less than 2% of the CPU time, and that
of the secure roaming support ranges from 0.2% to 5% of the CPU time depending on the
number of hosts and their motion. This justifies the feasibility of the security mechanisms.
1
1 INTRODUCTION
1.1 Problem statement
The research problem ishow to provide continuous connectivity for a mobile unit to
a network in which every node is moving. We investigate this problem in two network
environments shown in Figure 1.1: (a) mobile ad hoc networks that have no centralized
control; (b) large scale heterogeneous wireless networks with control in movable base
stations. The major challenges aredynamic topology, decentralized control, andlimited
bandwidth. We concentrate on research problems at the network layer.
The Internet
Satellite network
Cellular systems
Sensor network
Ad hoc network
Ad hoc networkWireless network withmovable base stations
Figure 1.1. Network environments
2
1.1.1 Adaptive routing in mobile ad hoc networks
A mobile ad hoc network (MONET) is a collection of mobile hosts that are deployed as
a multi-hop wireless network without the aid of any preexisting infrastructure or central-
ized administration. It relies on hosts cooperation to maintain network connectivity and
functionality. The salient characteristics of ad hoc networks, including highly dynamic
topologies, low bandwidth, energy-constrained operations, and limited computation capa-
bility, make the design of routing protocols a challenging problem. The protocols must
be capable of keeping up with the drastically and unpredictably changing network topol-
ogy, with minimized message exchanges, in a fully distributed way. Most proposed ad
hoc routing protocols, such as destination-sequenced distance vector (DSDV) [1], ad-hoc
on-demand distance vector (AODV) [2], and dynamic source routing (DSR) [3], adopt the
content of routing information from the Internet protocols and react to topology changes.
Research is needed to develop a protocol that is able to adapt to various network conditions
such as traffic load and congestion. This requires the following:
• Identifying the network parameters that impact the performance of routing protocols
and evaluating their effects through experiments.
• Determining the appropriateness of on-demand and proactive approaches to main-
tain network connectivity, given specific routing requirements and network param-
eters.
This research provides a better understanding of the advantages and disadvantages of
different routing approaches in various network contexts that will lead to the development
of new adaptive routing protocols. It offers guidelines on identifying and avoiding the
performance bottleneck of routing protocols, and choosing proper parameters in future
simulation and analytic studies. Based on this research, a congestion avoidance routing
protocol is developed, which is capable of adapting to congestion, traffic load, and network
topology.
3
1.1.2 Large scale heterogeneous wireless networks
The mushrooming of heterogeneous wireless technologies and the need of robust and
efficient communication systems call for the ubiquitous and integrated wireless infrastruc-
ture. While the existing wireless networks have been individually studied, the integrated
wireless system brings new challenges in network management, protocol design, and per-
formance evaluation.
In a heterogeneous wireless network, there are base stations that have more resources
than mobile hosts in terms of energy supply, processing power, etc. These base stations
that have multiple wireless interfaces (each interface may use different wireless technol-
ogy) connect different wireless networks. Most existing heterogeneous wireless network
models assume base stations are stationary [4–6]. They are not able to adapt to dynamic
movement. We study the case where even base stations are moving. We refer this kind
of network aswireless network with movable base stations(WNMBS). The following
research problems have been investigated.
• How to organize the network in an efficient way so that the effect of motion on
topology is minimized without loss of network connectivity? How to minimize the
involvement of resource-poor mobile hosts?
• How to build efficient routing protocols for WNMBS? The protocols should be IP-
compliant and transparent to upper layer services such as TCP and UDP. It should
be capable of cooperating with various routing protocols used by different sub-nets.
• What cryptography mechanism should be used to secure communications? How to
authenticate a mobile host? How to maintain authentication when a host is roaming
among the system? How to handle agent failures when authentication is required?
In addition to the commercial 3G wireless system that provides different mobile ser-
vices, many existing and future military networks that consist of highly dynamic au-
tonomous topology segments require integration of heterogeneous wireless technologies.
This research have impacts on the development of a framework to seamlessly support
4
IP-compliant data services over heterogeneous wireless networks and new security mech-
anisms that fit into the mobile wireless world.
1.2 Thesis contributions
1.2.1 Mobile ad hoc networks
We investigate the performance issues of DSDV and AODV routing protocols for mo-
bile ad hoc networks. Four performance metrics are measured by varying the maximum
speed of mobile hosts, the number of connections, and the network size. The correlation
between the network topology change and the host mobility is investigated by using linear
regression analysis. The simulation results indicate that AODV outperforms DSDV in less
stressful situations, while DSDV is more scalable with respect to the network size. It is
observed that network congestion is the dominant reason for packet drop for both proto-
cols. We propose a new routing protocol, congestion-aware distance vector (CADV), to
address the congestion issues. CADV outperforms AODV in delivery ratio by about 5%,
while introduces less protocol load. The result demonstrates that integrating congestion
avoidance mechanisms with proactive routing protocols is a promising way to improve
performance.
We study the impact of congestion and mobility on the packet loss in various network
contexts. The results indicate that DSDV loses 10% to 20% more packets than AODV
for UDP traffic. For TCP traffic, the packet loss for DSDV is a half of that for AODV.
Mobility is the dominant cause for AODV, which is responsible for more than 60% of
total packet loss. For DSDV, more than 50% of total packet loss is congestion-related.
This work provides guidelines for the design of routing and flow control algorithms and
insights in choosing proper parameters in future simulation and analytic studies.
Congestion in ad hoc networks is a serious problem. Contention among neighbors
for the access to the shared media is the primary cause for the network congestion. We
consider congestion in the design of the routing protocols. The main thrust is to avoid
congestion by minimizing contentions for channel access. The intermediate delay (IMD)
5
is proposed as a routing metric. It enables routing protocols to select routes that bypass
mobile nodes in contention. IMD characterizes the impacts of channel contention, traffic
load, and the length of a route. The packet transmission procedure of the distributed
coordination function (DCF) in the IEEE 802.11 standard is analyzed and used as a study
case for evaluation and experimentation. An estimate of the transmission delay is derived
based on local information available at a node. The estimation takes the impact of active
traffic in the neighborhood into account without exchanging messages with neighbors.
The self-adjusting congestion avoidance (SAGA) routing protocol is designed with
IMD as the routing metric. The performance of SAGA is evaluated and compared with
that of AODV, DSDV, and dynamic source routing (DSR) protocols using simulation. Two
types of UDP traffic, constant bit rate traffic and traffic exhibiting long range dependency,
as well as the TCP traffic are considered. SAGA can sustain heavier traffic load and
offers higher peak throughput than AODV and DSR. The overhead of SAGA can be as
low as 10% as that of AODV and 12% as that of DSR. The average end-to-end delay of
SAGA is the lowest among the protocols. It is shown that considerations of congestion and
intermediate delay instead of hop count can enhance routing performance significantly.
1.2.2 Wireless networks with movable base stations
Wireless networks with movable base stations combine the advantages of mobile ad
hoc networks and wireless LAN to achieve both flexibility and scalability. We present the
hierarchical mobile wireless network (HMWN) to support movable base stations. HMWN
may be applied to ad hoc networks as well to build a virtual hierarchy. In such a system,
mobile hosts are organized into hierarchical groups. Four basic operations for setting
up and maintaining the network structure are grouping, registration, leaving, and mi-
gration. An efficient group membership management protocol is developed to support
mobile hosts roaming among different groups. The segmented membership-based group
routing (SMGR) protocol is proposed to take advantage of the hierarchical structure and
membership information. In this protocol, only local message exchanging is required
6
for maintaining network topology and routing information. Simulation-based experiment
demonstrates the scalability of the design in terms of protocol overheads.
Security, flexibility, and scalability are critical to the success of wireless communi-
cations. In a HMWN system, the group agents serve as a distributed trust entity. A se-
cure packet forwarding algorithm and an authentication and key exchange protocol are
developed to protect the network infrastructure. A roaming support mechanism and the
associated mutual authentication protocol are proposed to secure the foreign group and
the mobile host when it roams within the network. The computation overhead of secure
packet forwarding and roaming support algorithms is studied via experiments. The results
demonstrate that the computation overhead of secure packet forwarding is less than 2%
of the CPU time, and that of secure roaming support ranges from 0.2% to 5% of the CPU
time depending on the number of hosts and their motion.
1.3 Thesis organization
Chapter 2 briefly introduces the two ad hoc routing protocols, DSDV and AODV, that
are used in this research. The network simulator ns2 and the specifications of the physical
and MAC layers of the wireless networks are described. The random waypoint mobility
model is used to generate movements for mobile hosts. Its parameters and characteristics
are outlined.
Chapter 3 presents the study of ad hoc routing protocols. The correlation between
the network topology change and the host mobility is investigated. The results indicate
that the topology change may be a linear function of the maximum speed and the pause
time, respectively. DSDV and AODV are taken as the representatives of the proactive
and on-demand routing approaches in this study. The performance of these protocols are
evaluated by varying the maximum speed, the number of connections, and the number of
mobile hosts. The measurements include delivery ratio, average end-to-end delay, proto-
col overhead, and power consumption. Further investigation on DSDV is conducted by
7
reducing the broadcast interval and increasing the queue length. Based on the analysis of
the experimental study, the congestion-aware routing protocol CADV is proposed.
Chapter 4 studies the causes for packet loss in ad hoc networks. The causes are catego-
rized as congestion-related and mobility-related. The effects of congestion and mobility
on DSDV and AODV in various network contexts are explored. The results indicate that
mobility is the dominant cause for AODV, which is responsible for more than 60% of total
packet loss. For DSDV, more than 50% of total packet loss is congestion-related.
Based on the guidelines obtained from the experimental studies. The self-adjust con-
gestion avoidance (SAGA) routing protocol is presented in Chapter 5. The characteristics
of contention-based access to wireless channels and their impact on ad hoc routing are
discussed. The intermediate delay (IMD) is proposed as a routing metric and the ideas of
ad hoc routing based on IMD are illustrated. The delay estimation is critical in SAGA.
When a node has active traffic, statistical methods are used to evaluate the mean of the
delay. Otherwise, the underlying MAC protocol is analyzed and probability methods are
applied to compute the expectation of the delay. The packet transmission procedure of
the distributed coordination function (DCF) in the IEEE 802.11 standard is analyzed and
used as a study case for evaluation and experimentation. The performance of SAGA is
evaluated and compared with that of AODV, DSR, and DSDV protocols.
In Chapter 6, a hierarchical network architecture is proposed to support movable base
stations in heterogeneous wireless networks. The design considerations include the asym-
metric capacity and responsibility between base stations and mobile hosts, the coordinated
movement of hosts, and the localized traffic. Four basic operations, grouping, registration,
leaving, and migration, are defined for setting up and maintaining the network structure.
The details of the membership management scheme and the segmented membership-based
group routing protocol are presented. Experiments are conducted to study the protocol
overhead.
Chapter 7 presents mechanisms for securing wireless networks with movable base sta-
tions. The base stations serve as a distributed trust entity for key management and authen-
tication. A secure packet forwarding algorithm and an authentication and key exchange
8
protocol are developed to protect the network infrastructure. A roaming support mecha-
nism and the associated mutual authentication protocol are proposed to secure the foreign
group and the mobile host when it roams within the network. The computation overhead
of secure packet forwarding and roaming support algorithms is studied via experiments.
Chapter 8 concludes this dissertation and outlines directions for extending the research.
9
2 BACKGROUND
2.1 Destination-sequenced distance vector routing protocol (DSDV)
DSDV routing protocol is one of the first routing protocols designed specially for ad
hoc networks. It extends the basic Bellman-Ford mechanism by attaching a sequence
number, which is originated by the destination, to each distance. This destination se-
quence number is used to determine the “freshness” of a route. Routes with more recent
sequence numbers are preferred for making packet forwarding decisions by a host, but
not necessarily advertised to other hosts. For routes with the equal sequence number, the
one with the smallest distance metric is chosen. Each time a host sends an update to its
neighbors, its current sequence number is incremented and included in the update. The
sequence number is disseminated throughout a network via update messages. The DSDV
protocol requires each host to periodically advertise its own routing table to its neighbors.
Updates are transmitted immediately when significant new routing information is avail-
able. Routes received in broadcasts are used to update the routing table. The receiver adds
an increment to the metric of each received route before updating.
In DSDV, the broken link may be detected by the layer-2 protocol, or may be inferred
if no broadcast has been received from a former neighbor for a while (e.g., three periodic
update periods). A broken link is assigned a metric of∞ (i.e., a value greater than the
maximum allowed metric). When a broken link to a next hop is detected, the metric of
any route through that next hop is immediately assigned∞, and the sequence number
associated with it is incremented. Such modified routes are immediately broadcast in
a routing update packet. Handling broken links is the only situation when a sequence
number is generated by a host other than the destination. To distinguish this situation,
sequence numbers generated by the originating hosts are even numbers, while sequence
10
numbers generated to indicate the∞ metric are odd numbers. Anyreal sequence number
will supersede an∞ metric.
Two types of updates are defined in DSDV protocol. One, called “full dump”, carries
all the available routing information. The other, called “incremental”, carries only infor-
mation changed since the last full dump. Full dumps are generated relatively infrequently.
If the size of an incremental approaches the size of a packet, a full dump can be scheduled
so that the next incremental will be smaller.
Since all mobile hosts periodically advertise their routing information, a host can al-
most always locate every other host when it needs to send out a packet. Otherwise, the
packet is queued until the routing information is available. DSDV guarantees loop-free
paths to each destination [1].
2.2 Ad hoc on-demand distance vector routing protocol (AODV)
AODV routing protocol is also based upon distance vector, and uses destination se-
quence numbers to determine the freshness of routes. It operates in the on-demand fash-
ion, as opposed to the proactive way of the DSDV protocol. AODV requires hosts to
maintain only active routes. Anactive routeis a route used to forward at least one packet
within the pastactive timeoutperiod. When a host needs to reach a destination and does
not have an active route, it broadcasts a Route Request (RREQ), which is flooded in the
network. A route can be determined when RREQ is received either by the destination
itself or by an intermediate host with an active route to that destination. A Route Replay
(RREP) is unicast back to the originator of RREQ to establish the route. Each host that
receives RREQ caches a route back to the originator of the request, so that RREP can be
sent back. Every route expires after a predetermined period of time. Sending a packet via
a route will reset the associated expiry time.
Every host monitors the link status of next hops in active routes by listening for “Hello”
messages from its neighbors or for any suitable link layer notification (such as those pro-
vided by IEEE 802.11). When a link break in an active route is detected, a Route Error
11
(RERR) is sent back along the path to the source. All hosts on that path notice the loss
of the link. In order to report errors, every host maintains aprecursor listfor each route,
containing the neighbors that are likely to forward packets on this route.
To prevent unnecessary network-wide dissemination of route request messages, the
source may use anexpanding ring searchtechnique as an optimization. The search range
is controlled by the time-to-live (TTL) field in the IP header of the RREQ packet. The
search process is repeated with an incremented TTL (thus expanding the ring) until a
route is discovered.
Another optimization islocal repair. When a broken link in an active route is detected,
instead of sending back RERR, the host first tries to locally repair the link by broadcast-
ing RREQ for the destination. Although local repair is likely to increase the number of
deliverable data packets, it may result in increased delay as well.
2.3 Simulation environment
This research involves extensive experimental studies using ns2. ns2 is an event-driven
network simulator targeted at networking research. It is a widely used tool for simulating
inter-network topologies to test and evaluate various networking protocols. It supports
simulations of wireless networks and interconnecting wired and wireless networks.
In simulation, each mobile host uses an omni-directional antenna having unity gain.
The wireless interface works like the 914 MHz Lucent WaveLAN direct-sequence spread-
spectrum (DSSS) radio interface [7]. WaveLAN is modeled as a shared-media radio with
a nominal bit rate of 2 Mb/s, and a nominal radio range of 250m [8]. The IEEE 802.11
distributed coordination function (DCF) is used as the MAC layer protocol. A unicast data
packet destined to a neighbor is sent out after handshaking with request-to-send/clear-to-
send (RTS/CTS) exchanges and followed by an acknowledgement (ACK) frame. The
broadcast packets are simply sent out without handshake and acknowledgement. The
implementation uses carrier sense multiple access with collision avoidance (CSMA/CA).
12
The implementations of DSDV and AODV provided by ns2 are used in the studies.
They closely match the specifications [1] and [2]. The implementation of AODV enables
expanding ring search and local repair.
2.4 Mobility model
The random waypointmodel [9] is used to generate movements for mobile hosts. At
the beginning of a simulation, mobile hosts are randomly placed on a 1000m x 1000m
square field. Each host randomly chooses its destination in the field, and a moving speed
that ranges from 0 to the given maximum speed. All destinations and speeds are indepen-
dent and identically distributed. After a host reaches the destination, it waits for a speci-
fied time (i.e., pause time), and then repeats the above steps. According to this model, the
speed and direction of the next movement have no relation to those of the previous move-
ment. As indicated in [10], the pause time and the maximum speed have similar impacts
on the mobility with respect to link change or route change. Thus the mobility is varied
by changing the pause time or the maximum speed in the simulation.
13
3 STUDY OF AD HOC ROUTING PROTOCOLS
3.1 Introduction
3.1.1 Problem statement
The high mobility, low bandwidth, and limited computing capability characteristics of
mobile hosts make the design of ad hoc routing protocols challenging. The protocols must
be able to keep up with the drastically and unpredictably changing network topology, with
minimized message exchanges, in a computation efficient way.
The routing protocols may be categorized asproactive, on-demand, andhybrid, ac-
cording to the way in which the mobile hosts exchange routing information. The proactive
protocols, such as DSDV [1] and source tree adaptive routing (STAR) [11,12], periodically
disseminate routing information among all the hosts in the network, so that every host has
the up-to-date information for all possible routes. On-demand routing protocols, such as
AODV [2] and dynamic source routing (DSR) [3], operate on a need basis, discover and
maintain only active routes that are currently used for delivering data packets. Hybrid
routing protocols, such as zone routing protocol (ZRP) [13, 14] and Core Extraction Dis-
tributed Ad Hoc Routing (CEDAR) [15], maintain a virtual routing infrastructure, apply
proactive routing mechanisms in certain regions of a network and on-demand routing in
the rest of the network.
An ad hoc routing protocol tends to be well-suited for some network contexts, yet less
suited for the others [16]. A better understanding of the advantages and disadvantages of
different routing approaches in various network contexts will serve as a cornerstone for
the development of new adaptive routing protocols. However, ad hoc networks are too
complex to allow analytical study for explicit performance expressions. We use the means
14
of simulation to evaluate the routing approaches numerically and gather data to estimate
their characteristics.
We study the performance of DSDV and AODV in a wide range of network contexts
with varied network size, mobility, and traffic load. Both protocols utilize distance vector
coupled with destination sequence number, and choose routes in the same manner. They
are differentiated by the way in which they operate (i.e., proactive versus on-demand).
Studying these two protocols gives insights into the differences between proactive and
on-demand approaches. This analysis provides guidelines to improve these two specific
protocols as well.
3.1.2 Our contributions
The linear dependence between network topology change and host mobility is inves-
tigated by using statistical analysis. The suitable network contexts for DSDV and AODV
are identified. We discover that AODV introduces 1.5 to 5 times protocol load as DSDV
does, which contradicts the motivation for the on-demand approach. The major causes for
packet drop are investigated by exploring packet traces. We argue that DSDV is plagued by
network congestion. Based upon the idea of integrating congestion avoidance mechanisms
with proactive routing protocols to improve routing performance, we propose congestion-
aware distance vector (CADV) routing protocol. The preliminary study of CADV shows
positive results. To our knowledge, it is the first research effort to take the power con-
sumption as a routing performance metric.
3.2 Related work
Several simulation-based performance comparisons have been done for ad hoc routing
protocols in the recent years. Das et al. evaluate performance of ad hoc routing protocols
based on the number of conversations per mobile node using Maryland Routing Simulator
(MaRS) [17]. The performance comparison of two on-demand routing protocols DSR and
AODV is presented in [8], using ns2 (network simulator) [18] for the simulation. The
15
pause time and the offered traffic load are taken as parameters. In [19], GloMoSim [20] is
used for the performance study of the STAR, AODV, and DSR routing protocols, taking
the pause time as the parameter. The authors point out that simulating the same protocol
in different simulators may produce differences in the results. The performance of two
location-based routing protocols for ad hoc networks is investigated by using ns2 and the
effect of average moving speed in different scenarios is presented in [21]. An adaptive
distance vector routing algorithm is proposed in [22], and its performance, compared with
AODV and DSR, is studied. The offered traffic load and the simulation time are the input
parameters.
Our work is to comprehensively investigate the characteristics of proactive and on-
demand approaches by studying DSDV and AODV. In addition to identifying the suitable
network contexts for each approach, we explore the causes for performance degradation.
Based on the investigation, a new distance vector based routing protocol is proposed.
The rest of the chapter is organized as follows. In Section 3.3, the correlation between
topology change and mobility is investigated. Section 3.4 describes the simulation envi-
ronment, including the mobility, traffic, energy models, and performance metrics. Section
3.5 presents the experiment results and analysis. Improvements of DSDV are discussed
in Section 3.6. Section 3.7 introduces the proposed CADV routing protocol and presents
preliminary results of performance comparison of CADV, DSDV, and AODV. Section 3.8
concludes this chapter.
3.3 Correlation between topology change and mobility
The performance of a routing protocol is effected by the rate of topology change (i.e.,
the speed at which a network’s topology is changing). The topology change can be repre-
sented as link change or route change. It is difficult to control the either of them directly
in simulations. Our study demonstrates that:
• The link change and route change can be perfectly fitted into linear functions of the
maximum speed when the pause time is 10 seconds.
16
0 5 10 15 20 250
500
1000
1500
2000
2500
3000
3500
Link
Cha
nges
in S
imul
atio
n
Max Speed (m/s)
Linear Fitting: Y = 142.3839*x + 474.4091
|t| = |b1|/σ*sqrt(S
xx) = 24.1445
Sample PointsFitting CurveConfidence Interval of 95%
(a)
0 5 10 15 20 250
0.5
1
1.5
2
2.5
3
x 104
Rou
te C
hang
es in
Sim
ulat
ion
Max Speed (m/s)
Linear Fitting: Y = 1302.5252*x + 4668.6727|t| = |a|/σ*sqrt(S
xx) = 21.1927
(b)
0 100 200 300 400 5000
100
200
300
400
500
600
700
800
Link
Cha
nges
in S
imul
atio
n
Pause Time(s)
Linear Fitting: Y = −1.2172*x + 834.7333|t| = |a|/σ*sqrt(S
xx) = 9.1826
(c)
0 100 200 300 400 5000
1000
2000
3000
4000
5000
6000
7000
8000
Rou
te C
hang
es in
Sim
ulat
ion
Pause Time(s)
Linear Fitting: Y = −10.339*x + 8580.6762|t| = |a|/σ*sqrt(S
xx) = 8.0857
(d)
Figure 3.1. Topology change vs. mobility
• The link change and route change can be perfectly fitted into linear functions of the
pause time when the maximum speed is 4 m/s.
Thus, the topology change can be indirectly controlled by varying mobility.
As shown in Figure 3.1a and 3.1b, the maximum speed is treated as the predictor
variable, and link changes and route changes as the response variables (with the pause
time to be 10 seconds). The fitting curve is obtained by using linear regression with least
squares [23].
�Y = b0 + b1X
b1 =
∑ni=1 XiYi − nXY
∑ni=1 Xi
2 − nX2
17
b0 = Y − b1X
If we assume that the variations of the sample points about the line are normal, we can test
the null hypothesis:
H0 : b1 = 0
using thet-test[23].
t =b1
√
ni=1 (Xi − X)2�
σ�σ2 =
ni=1 (Yi −�Yi)
2
n − 2
For the link changes versus the maximum speed,|t| = 24.1445. For the route changes
versus the maximum speed,|t| = 21.1927. Both of them exceed the appropriate critical
value of t0.995(10) = 3.1691 (because 12 sample points are used for the linear regression,
the degree of freedom is 10 = 12 - 2). Thus the hypothesisH0 that the linear relationships
between the link changes and the maximum speed, the route changes and the maximum
speed do not exist is rejected with 99% confidence. The dotted lines in Figure 3.1 indicate
the confidence interval of 95%. In plain words, the values of the link changes and the
route changes lie within the specified intervals, respectively, and the statement is made
with 95% confidence.
Figure 3.1c and 3.1d show the linear regressions of the link change versus the pause
time and the route change versus the pause time.H0 hypothesis is also verified with t-
test. Because only 6 sample points are used, the degree of freedom is 4. t0.995(4) = 4.604,
while the observed|t| is 9.1826 and 8.0857 respectively. ThusH0 is rejected with 99%
confidence as well. The dotted lines in figure 3.1c and 3.1d show the confidence intervals
of 95%.
3.4 Simulation settings and performance metrics
The constant bit rate (CBR) traffic is used in the simulation. Each connection is speci-
fied as a randomly chosen source-destination (S-D) pair. The packet sizes are fixed as 5121The percentage points for the t-distribution are obtained from [23], using the two-tailed table.
18
Table 3.1Power requirements
State Documented RequirementsMeasured
suspended 0.00 W 0.00 W
receiving 1.48 W 1.52 W
transmitting 3.00 W 3.10 W
bytes. The packet sending rate is 4 packets per second. Each connection starts at a time
randomly chosen from 0 to 100 seconds.
Every host has an initial energy level at the beginning of a simulation. For every trans-
mission and reception of packets, the energy level is decremented by a specified value,
which represents the energy usage for transmitting and receiving. When the energy level
goes down to zero, no more packets can be received or transmitted by the host. According
to the manufacturer specifications [7], the power requirements of the WaveLAN card are
shown in Table 3.1, column 2. Column 3 shows the actual power requirements measured
in [24], without any power management mechanism. In the simulations, we use the values
in column 3. We let the initial energy of each host to be 4000 joules so that the energy
level does not reach zero in the simulation period.
The following four quantitative metrics are used to assess the performance:
• Delivery Ratio: The ratio of the data delivered to the destinations (i.e., throughput)
to the data sent out by the sources.
• Average End-to-end Delay: The average time it takes for a packet to reach the des-
tination. It includes all possible delays in the source and each intermediate host,
caused by routing discovery, queueing at the interface queue, transmission at the
MAC layer, etc. Only successfully delivered packets are counted.
• Protocol Overhead: The routing load per unit data successfully delivered to the des-
tination. The routing load is measured as the number of protocol messages trans-
19
mitted hop-wise (i.e., the transmission on each hop is counted once). A unit data
can be a byte or a packet.
• Power Consumption: The total consumed energy divided by the number of delivered
packets. We measure the power consumption because it is one of the precious com-
modities in mobile communications. Wireless devices may consume over 50% of
total system power for current handhold computers, and up to 10% for high-end lap-
tops [24]. This poses challenging demands on the design of power-efficient routing
protocols.
In the simulation, five scenarios are generated using the random waypoint model for
each experiment, and the average values are used for analysis.
3.5 Results and analysis
To comprehensively measure the performance of a protocol, various network contexts
are considered. The following parameters are varied in the simulation.
• Host mobilityis determined by the maximum speed (with 10 seconds pause time).
• Traffic loadis the number of the CBR connections.
• Network sizeis measured as the number of mobile hosts. Since the simulation field
is fixed, the network size also measures the density of mobile hosts.
3.5.1 Varying maximum speed
This set of experiments study the impact of mobility on the performance metrics. The
number of mobile hosts and the number of connections are both 30. The maximum speed
ranges over{4, 8, 12, 16, 20, 24} m/s.
20
0 5 10 15 20 250
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Del
iver
y R
atio
Maximum Speed (m/s)
AODVDSDV
(a)
0 5 10 15 20 250
0.5
1
1.5
2
2.5
Ave
rage
End
−to
−en
d D
elay
(s)
Maximum Speed (m/s)(b)
0 5 10 15 20 250
0.05
0.1
0.15
0.2
0.25
Pro
toco
l Ove
rhea
d
Maximum Speed (m/s)(c)
0 5 10 15 20 250
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
x 10−3
Pow
er C
onsu
mpt
ion
Maximum Speed (m/s)(d)
Figure 3.2. Varying maximum speed
As Figure 3.2a shows, the packet delivery ratios for both protocols are less than 50%2.
When mobility is low (i.e., the maximum speed is 4 m/s), AODV delivers about 43% of
total packets, while DSDV delivers about 34%. As the mobility increases, the delivery
ratios of both protocols drop gradually, but DSDV has a little bigger drop.
It is interesting that DSDV has a higher delay than AODV does in all cases, which
seems to contradict to the advantage of the proactive approach. It results from the imple-
mentations of the protocols. Although both implementations apply the drop-tail approach
2The implementation of IEEE 802.11 has been revised in ns2 since version 2.1b9. The default wirelessbandwidth is set to 1 Mb/s. It, however, does not affect the performance comparison in this chapter, becauseit has same impact on different routing protocols.
21
for packet queues, AODV poses a limit on the time a packet can be queued, which cur-
rently is 30 seconds. Thus the delay of any received packet is bounded. DSDV keeps
packets in queues no matter how long they have stayed. It delivers the older packets rather
than the younger ones, and therefore increases the average delay.
Because a DSDV protocol packet contains many routes, while an AODV protocol
packet contains at most one route (e.g., RREQ), we compare the byte-wise protocol over-
head. DSDV introduces a significantly (3-4 times) lower protocol overhead than AODV
(Figure 3.2c). The bad performance of AODV results from the following factors:
• Each host discovers routes individually.
• Unicasting RREP to the originator of the RREQ prevents valuable routing informa-
tion from being propagated to other hosts.
• AODV treats network topology as a directed graph. It might need to discover two
different directions for the same path twice due to a short reverse route lifetime.
As illustrated in Figure 3.2d, the power consumptions for both protocols are rather
stable. Although DSDV introduces a much lower protocol overhead, it consumes more
power. AODV “wins” in the way it handles link breaks. When a broken link of a route
is detected, a route error (RERR) packet is sent to the source. Every host along the path
notices the broken link immediately, and drops or queues packets locally. DSDV treats a
broken link as a significant routing information and triggers a routing update. There is a
minimum time interval between two triggered updates. The information about a broken
link is delayed at each host. In the meantime, those hosts that have not received this
information keep sending packets that will be dropped eventually to their next hops. A
remarkable amount of power is consumed unnecessarily.
22
10 20 30 40 50 600
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Del
iver
y R
atio
Number of Connections
AODVDSDV
(a)
10 20 30 40 50 600
0.5
1
1.5
2
2.5
3
3.5
4
Ave
rage
End
−to
−en
d D
elay
(s)
Number of Connections(b)
10 20 30 40 50 600
0.05
0.1
0.15
0.2
0.25
Pro
toco
l Ove
rhea
d
Number of Connections(c)
10 20 30 40 50 600
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6x 10
−3
Pow
er C
onsu
mpt
ion
Number of Connections(d)
Figure 3.3. Varying number of connections
3.5.2 Varying number of connections
The next set of experiments demonstrate the effect of the traffic load. The number of
mobile hosts is 30, the maximum speed is 4 m/s, and the pause time is 10 seconds. The
number of connections varies from 10 to 80, increasing by 10 each time.
The delivery ratio of AODV (Figure 3.3a) drops dramatically from more than 90%
to about 28% when the number of connections increases from 10 to 50, while that of
DSDV drops from about 80% to about 20%. For more than 50 connections, the ratios of
both DSDV and AODV drop more gradually because the network has already been fully
loaded.
23
As Figure 3.3b shows, for 10 connections, DSDV and AODV have similar delay. The
delays for both protocols increase rapidly with the number of connections (from about 0.1
second to 3 and 2.5 seconds for 40 connections, respectively). After the number of con-
nections reaches 40, the delay of AODV grows gradually, while that of DSDV increases
almost as fast as before.
For DSDV, the number of protocol packets is determined mostly by the network size
and mobility. The protocol overhead stays fairly stable at 0.06 with an increasing number
of connections (figure 3.3c). The protocol overhead of AODV increases sharply as the
number of connections increases. AODV performs better than DSDV at 10 connections.
At 80 connections, the protocol overhead for AODV is about 4 times higher than for
DSDV.
As shown in Figure 3.3d, DSDV consumes more power than AODV does except for
10 connections. The power consumptions for both protocols increases gradually from 10
connections to 80 connections (the increase is about 50% for DSDV, and about 25% for
AODV).
3.5.3 Dropped packets
Since the delivery ratio drops dramatically with an increase in traffic load, we are
interested in investigating the reasons for packet drop. We check this by studying the ns2
trace files.
Figure 3.4 shows the number of packets dropped for four reasons. A packet is dropped
due to congestion if the packet buffer at MAC layer is full when it arrives. When a collision
is detected, CSMA/CA does a exponential backoff, which increases the delay for sending
the packet. It makes the packet buffer to be full quickly.
For DSDV, no packet is dropped due to “no route” to the destination. It is guaranteed
by the design the protocol. For AODV, the number of packets dropped due to “no route”
increases from 2000 to 10000, as shown in Figure 3.4a.
24
10 20 30 40 50 600
1000
2000
3000
4000
5000
6000
7000
8000
9000
Number of Connections
Dro
pped
For
No
Rou
te
AODVDSDV
(a)
10 20 30 40 50 600
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
x 104
Number of Connections
Dro
pped
For
MA
C C
allb
ack
(b)
10 20 30 40 50 600
0.5
1
1.5
2
2.5
x 104
Number of Connections
Dro
pped
For
Que
ue F
ull
(c)
10 20 30 40 50 600
2
4
6
8
10
12
x 104
Number of Connections
Dro
pped
For
Oth
er R
easo
ns
(d)
Figure 3.4. Dropped packets
As Figure 3.4b and 3.4c show, for 10 connections, AODV almost does not drop pack-
ets due to a MAC callback (i.e., the next hop is not a neighbor now), or queue being full.
However, the number of packets dropped for AODV increases with the number of connec-
tions at a rate higher than DSDV. DSDV drops fewer packets than AODV for the above
two reasons in most cases.
From Figure 3.4, we can calculate that more than half of the dropped packets result
from congestion. DSDV performs better for the first three reasons, but worse than AODV
for avoiding congestion. Although both DSDV and AODV do not utilize any congestion
control or avoidance mechanism to balance traffic load, AODV in fact distributes the data
25
traffic more evenly in the network. AODV tries to build the shortest route when it origi-
nates a request, but it keeps the route as long as it does not break, even if a shorter route is
available at a later time. In contrast, DSDV tends to always send packets via the shortest
routes. Forwarding packets through the shortest routes will likely push traffic to several
heavily burdened hosts and congest the network.
3.5.4 Varying number of mobile hosts
The last set of experiments investigate the effect of the network size. All hosts move
randomly at the maximum speed of 4 m/s. The pause time between two movements is
10 seconds. The number of mobile hosts increases from 20 to 70 by 10s. The number of
connections is equal to the number of hosts.
The delivery ratio of AODV decreases faster than that of DSDV with the number of
mobile hosts (Figure 3.5a). AODV has a better performance in a sparser network (fewer
than 40 hosts), and worse performance in a denser one. Figure 3.5b indicates that AODV
outperforms DSDV in terms of end-to-end delay.
DSDV and AODV have similar protocol overhead for 20 mobile hosts. Both of them
introduce more overhead as the number of hosts increases, with the overhead for AODV
growing faster than for DSDV (Figure 3.5c).
Both DSDV and AODV have similar power consumption in a sparse network (Figure
3.5d). For DSDV, the increase of power consumption is nearly linear with the host number.
The power consumption for AODV increases faster than for DSDV. For 70 hosts, AODV
consumes 33% more energy than DSDV does per 1k-byte delivered data.
From the results provided in Figure 3.5, we can tell that DSDV is more scalable with
respect to the number of hosts. It seems that 40 hosts per square kilometer is the turning
point. For more than 40 hosts, DSDV equals or outperforms AODV for all metrics (the
average delay is an exception that should not be considered).
26
20 30 40 50 60 700
0.1
0.2
0.3
0.4
0.5
0.6
Del
iver
y R
atio
Number of Mobile Hosts
AODVDSDV
(a)
20 30 40 50 60 700
1
2
3
4
5
6
7
8
9
Ave
rage
End
−to
−en
d D
elay
(s)
Number of Mobile Hosts(b)
20 30 40 50 60 700
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Pro
toco
l Ove
rhea
d
Number of Mobile Hosts(c)
20 30 40 50 60 700
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
x 10−3
Pow
er C
onsu
mpt
ion
Number of Mobile Hosts(d)
Figure 3.5. Varying number of mobile hosts
3.6 Further discussion about DSDV
3.6.1 Reduce broadcast interval of DSDV
The time interval between broadcasting routing information is one of the most impor-
tant parameters of DSDV [1]. As shown in figure 3.4, in total about5.5 ∗ 104 packets are
dropped for 80 connections due to either a MAC callback or a full queue, which means that
the outgoing links are broken or the routes are not established timely. Some of these situa-
tions could be avoided by broadcasting routing information more frequently, at the cost of
27
10 20 30 40 50 60 70 800.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Del
iver
y R
atio
Number of Connections
DSDV (default)DSDV (QLen. 64)DSDV (Update 8s)
(a)
10 20 30 40 50 60 70 800
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Ave
rage
End
−to
−en
d D
elay
(s)
Number of Connections(b)
10 20 30 40 50 60 70 800.05
0.06
0.07
0.08
0.09
0.1
0.11
0.12
0.13
Pro
toco
l Ove
rhea
d
Number of Connections(c)
10 20 30 40 50 60 70 801.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8x 10
−3
Pow
er C
onsu
mpt
ion
Number of Connections(d)
Figure 3.6. Performance comparison of different DSDV implementations
a higher protocol overhead. The questions are: How much improvement of performance
can be obtained? How much will it cost?
We reduced the broadcast time interval from 15 seconds to 8 seconds, and rerun the set
of experiments described in Section 3.5.2, using the same settings, parameters, scenarios,
and connections.
Figure 3.6a (the “Update 8s” curve) shows that the throughput increases about 10%
for less stressful cases (i.e., for fewer than 50 connection). The average delay is almost
the same (Figure 3.6b). The protocol overhead doubles as we expect (Figure 3.6c). The
28
power consumption slightly decreases, because packets are dropped earlier as we explain
in Section 3.5.1.
3.6.2 Increase the queue length of DSDV
Figure 3.4c shows that about1.5 ∗ 104 packets are dropped due to a full queue. Since
the queue length for DSDV is only 5, much smaller than that for AODV, it is natural to
ask this question: Will a longer queue increase the throughput of DSDV?
We set the queue length to 64 and rerun the set of experiments again. The results are
shown in Figure 3.6 (the “QLen. 64” curve). The performance metrics are almost the
same as those measured for the original DSDV implementation. Thus, the longer queue
does not help in improving performance of DSDV.
3.7 Congestion-aware routing protocol – CADV
Although the published result [25] showed that on-demand protocols outperform proac-
tive protocols and are better suited for mobile ad hoc networks, the proactive protocols
have the following advantages.
• Better support for Quality of Service (QoS):Proactive protocols timely propagate
network conditions (available bandwidth, delay, etc.) throughout the system, so that
appropriate QoS decisions, including admission control, traffic shaping, and route
choosing, can be made.
• Better support for anomaly detection:Proactive protocols constantly exchange the
network topology information. It enables real-time detection and reaction to ma-
licious behaviors and attacks such as the false distance vector attack and the false
destination sequence attack [26,27].
As shown in Section 3.5.4, DSDV performs better than AODV in denser networks, which
demonstrates potential scalability of the proactive approach with respect to the number of
mobile hosts. Figure 3.4 reveals that this approach is plagued by congestion, the dominant
29
reason of performance decrease. To address the congestion issues, we propose a new
proactive distance vector based ad hoc routing protocol called congestion-aware distance
vector (CADV).
3.7.1 Overview
A mobile host in an ad hoc network can be viewed as a single server queueing sys-
tem. The delay of sending a packet is positively correlated with congestion. In CADV,
each routing entry is associated with anexpected delay, which measures congestion at
the next hop. Every host estimates the expected delay based on the mean of delay for
all data packets sent in a past short period of time. Currently, the length of the period is
equal to the interval between two periodical updates. The expected delay is computed as
E[D] =∑
Di
nL, wheren is the number of sent packets andL is the length of MAC layer
packet queue.E[D] estimates the time a newly arrived packet has to wait before it is sent
out. When a host broadcasts an update to neighbors, it specifies the delay it may introduce.
A routing decision is made based on the distance to the destination as well as the expected
delay at the next hop. CADV tries to balance traffic and avoid congestion by giving pri-
ority to a route having low expected delay. For example, hosts A and B both advertise a
route to the destination. If the expected delay at host A is significantly less than that at
host B, A will be chosen as the next hop (given B is not A’s next hop), even if the route
via A is one hop longer than the one via B. When making routing decisions, a function
f(E[D], distance) is used to evaluate the value of a route. Various routing policies can be
implemented by replacing this function.
A CADV routing module consists of three components: (a)Traffic Monitor monitors
traffic going out through the link layer. It keeps track of the average delay for sending one
data packet in recent period of time. The time period is specified by the route maintenance
component. (b)Traffic Controldetermines which packet is the next to send or drop, and
reschedules packets if needed. It supports a drop tail FIFO queue and provides functional-
ity to re-queue packets. (c)Route Maintenanceis the core component. Its functionalities
30
10 20 30 40 50 600.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Number of Connections
Del
iver
y R
atio
AODVDSDVCADV
(a)
10 20 30 40 50 600
1
2
3
4
5
6
Ave
rage
End
−to
−en
d D
elay
(s)
Number of Connections(b)
10 20 30 40 50 600
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Number of Connections
Pro
toco
l Ove
ahea
d
(c)
10 20 30 40 50 600
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2x 10
−3
Pow
er C
onsu
mpt
ion
Number of Connections(d)
Figure 3.7. Comparison of three protocols
include exchanging information with neighbors, evaluating and maintaining routes, man-
aging the traffic monitor and traffic control components.
3.7.2 Preliminary results
A preliminary study is conducted to investigate the performance of CADV with the
number of connections. The maximum speed is 4 m/s and the number of mobile hosts
is 30. Figure 3.7 illustrates the performance comparison of CADV, DSDV, and AODV.
AODV performs better than CADV only for 10 connections, where congestion is not likely
to occur. For other cases, as shown in Figure 3.7a, CADV outperforms AODV by about
31
5% in terms of packet delivery ratio. The tradeoff for the improvement is shown in Figure
3.7c. CADV introduces about 2.5 times protocol overhead as DSDV does. However,
the protocol overhead is still lower than that introduced by AODV when the number of
connections is greater than 10. CADV introduces higher end-to-end delay than AODV
and DSDV when the number of connections is greater than 10 (figure 3.7b), because it
may choose longer route to forward packets. The delay is rather stable with the increase
of the number of connections. Figure 3.7d shows that CADV consumes less power. It
results from packet rescheduling done by the traffic control component. When a neighbor
becomes unreachable, all packets in the MAC layer packet buffer whose next hop is that
neighbor will be rescheduled. This mechanism saves power by preventing a host from
sending unnecessary Request-To-Send (RTS) messages.
3.8 Conclusion
Conclusion 1:For the movements of mobile hosts generated by the random waypoint
model, with a very high probability, the link change and route change are, linear functions
of the maximum speed, and linear functions of the pause time, respectively. The maximum
speed does not affect much the performance of DSDV and AODV at the range from 4 m/s
to 24 m/s.
Conclusion 2:In less stressful situations, AODV outperforms DSDV for all metrics
except for normalized protocol load. DSDV performs better than AODV does in denser
networks with a higher traffic load. In general, we can state: (1) The protocol load for the
proactive routing protocols (such as DSDV) grows as the number of hosts increases, while
that of the on-demand routing protocols (such as AODV) increases with the number of
source-destination (S-D) pairs. The proactive approach performs better when the number
of S-D pairs is close to the number of hosts. (2) The on-demand approach consumes
less power, because it propagates the link break information faster, thus it avoids sending
packets that are dropped eventually. (3) Network congestion is the dominant reason for
packet drop for both proactive and on-demand approaches.
32
Conclusion 3: The preliminary study of CADV routing protocol demonstrates that
the performance of proactive routing protocols can be improved by integrating with con-
gestion avoidance mechanisms. Currently, only delay at the next hop and distance to the
destination are considered when making routing decisions. We are working towards a
complete version of CADV that takes advantage of other information such as available
queue length, delay on a path, etc. A comprehensive study will be conducted to investi-
gate how different congestion predication and load balancing mechanisms can cooperate
with CADV to reduce congestion in ad hoc networks.
33
4 PACKET LOSS IN AD HOC NETWORKS
4.1 Introduction
Throughput is generally accepted as one of the most important metrics to evaluate the
performance of a routing protocol. Several simulation-based performance comparisons
have been done for ad hoc routing protocols in the recent years. S.R. Das et al. evaluate
performance of ad hoc routing protocols based on the number of conversations per mobile
node [17]. The performance comparison of two on-demand routing protocols: dynamic
source routing (DSR) [3] and AODV [2] is presented in [8]. The performance of two
location-based routing protocols for ad hoc networks is investigated in [21]. An adaptive
distance vector routing algorithm is proposed in [22], and its performance, compared with
AODV and DSR, is studied. Although various throughput results in different network
contexts have been obtained, the causes for throughput variation in ad hoc networks have
not been deeply understood. Packet loss is one thrust to study throughput, since throughput
is determined by how many packets have been sent and how many packets have lost.
Packet loss in wired network has been investigated. For example, a single server
queueing system with a finite buffer capacity is used to analyze packet loss processes
in high-speed networks in [28]. The end-to-end packet delay and loss behaviors in the
Internet are studied using the UDP echo tool in [29]. These work target at the packet loss
due to buffer overflow (congestion), which is the major loss in wired networks.
Packet loss problem is much more complicated in mobile ad hoc networks, because
wireless links are subject to transmission errors and the network topology changes dynam-
ically. A packet may lose due to transmission errors, no route to the destination, broken
links, congestions, etc. The effects of these causes are tightly associated with the net-
work context (e.g., host mobility, number of connections, traffic load, etc.). Even building
an approximate model to analytically evaluate packet loss is difficult. We investigate the
34
problem via simulations. Data is gathered from more than 1000 individual experiments to
estimate the desired true characteristics of packet loss in ad hoc networks.
In mobile ad hoc networks, wireless link transmission errors, mobility, and congestion
are major causes for packet loss. Packet loss due to transmission errors is affected by
the physical condition of the channel, the terrain where networks are deployed, etc. They
can not be eliminated or reduced by improving the routing protocols. This chapter only
addresses congestion-related and mobility-related packet loss. Congestion in a network
occurs whenever the demands exceed the maximum capacity of a communication link,
especially when multiple hosts try to access a shared media simultaneously. Mobility may
cause packet loss in different ways. A packet may be dropped at the source if a route to the
destination is not available, or the buffer that stores pending packets is full. It may also be
dropped at an intermediate host if the link to the next hop has broken. We study the effect
of congestion and mobility on packet loss in various network contexts. AODV and DSDV
are chosen as representatives of on-demand and proactive routing protocols respectively.
This work can benefit the design of routing and flow control algorithms, the dimen-
sioning of buffers, identifying and avoiding the performance bottleneck of current routing
protocols, and choosing proper parameters in future simulation and analytic studies.
The rest of the chapter is organized as follows. Section 4.2 introduces the related
work. The simulation settings, including traffic, routing protocols, congestion-related and
mobility-related packet loss, are discussed in Section 4.3. Section 4.4 presents two sets
of experiments and the results. The relations between the shortest path and congestion is
discussed in section 4.5. Section 4.6 concludes the chapter.
4.2 Related work
There has been some recent work on addressing packet loss issues in wireless net-
works. S. Biaz and N.H. Vaidya investigate the ability of three loss predictors to distin-
guish congestion losses from wireless transmission losses [30]. They use a wireless link
with transmission loss raterw in the simulations. F. Anjum and L. Tassiulas analytically
35
study the performance of different TCP algorithms over a wireless channel with correlated
packet losses [31]. A simple two-state Markov chain is used to model the correlated fading
channel. T.V. Lakshman et al. also analyze the impact of random packet loss at a wireless
link on the performance of TCP/IP in [32]. They indicate that bidirectional congestion
increases TCP’s sensitivity to loss. These efforts assume transmission losses on a single
wireless link follow a simple model and focus on how losses effect the performance of
TCP.
Even if wireless transmission is loss-free, packet loss still exists in ad hoc networks.
Our work is to understand the major causes for packet loss and to capture its characteris-
tics.
4.3 Simulation settings
4.3.1 Traffic
To investigate the impact of traffic load and congestion control mechanisms on packet
losses, both unresponsive traffic and responsive traffic are studied.
• Unresponsive trafficonly consists of UDP connections, each of which is specified
as a source-destination (S-D) pair. Every source is associated with a constant bit
rate (CBR) traffic generator, which sends out packets at the given rate. The source
of each S-D pair is randomly chosen from all hosts, and the destination is randomly
chosen from all hosts other than the source. All S-D pairs are mutually independent.
The packet size is fixed at 512 bytes. The start time of each connection is uniformly
distributed between 0 to 100 seconds.
• Responsive trafficis comprised of TCP connections. Each connection has a Tahoe
TCP 1 sender and a TCPSink receiver. The sender window size is decreased by
half when packet losses are detected. The retransmission starts from the first lost
packet. Tahoe TCP enters the slow start when an ACK for a new packet is received.
1The TCP performs congestion control and round-trip-time estimation in a way similar to the version ofTCP released with the 4.3BSD Tahoe UNIX system from UC Berkeley, so it is called Tahoe TCP.
36
Table 4.1Packet loss at MAC and network layers
Mobility-related Congestion-related
MAC Layer√ √
Network Layer√
All TCP packets have the same size of 512 bytes. The initial sender window size
is 1 and the maximum bound on the window size is 32. TCPSink is responsible
for returning ACKs to the sender. It generates one ACK per packet received. The
ACK packet size is 40. The data of each connection is generated by an attached FTP
application, which simulates a bulk data transfer. Every FTP application starts at a
time randomly chosen from 0 to 100 seconds.
4.3.2 Differentiated packet losses
Packet loss is measured at all mobile hosts. Every host monitors the networking layer
and the MAC layer for all kinds of packet losses. The layers of the protocol stack and the
modules that are responsible for mobility-related and congestion-related packet loss are
shown in Table 4.1.
Mobility-related packet loss may occur at both the network layer and the MAC layer.
When a packet arrives at the network layer, the routing protocol forwards the packet if a
valid route to the destination is known. Otherwise, the packet is buffered until a route is
available. A packet is dropped in two cases:
• The buffer is full when the packet needs to be buffered.
• The time that the packet has been buffered exceeds the limit. (The AODV imple-
mentation in ns2 poses a 30-second limit on the time a packet can be buffered. The
DSDV implementation does not have a limit.)
37
The MAC layer mobility-related packet loss occurs when the next hop of a packet is out of
range at the moment the packet is transmitted. The reason is that the routing information is
obsoleted. It occurs frequently in a high mobility network than in a low mobility network.
Congestion-related packet loss only occurs at the MAC layer. Because CSMA/CA is
used in the simulation, a packet may be dropped due to congestion for two reasons:
• The wireless channel is so busy that the number ofbackoffprocedures exceeds the
limit.
• The channel is associated with a queue that buffers all the packets waiting to be
transmitted. A packet is dropped if the queue is full when it arrives.
4.4 Experiments
A series of experiments have been conducted to investigate the mobility-related and
congestion-related packet losses in different network contexts. The network configuration
for the experiments is a 1000m x 1000m square field with 30 hosts. The buffer size is 64-
packet for each route and the MAC layer. Each data point in the result figures represents an
average of 5 runs with identical traffic but different mobility scenarios, which are randomly
generated with the same parameters (i.e., same maximum speed and pause time). Every
experiment runs for at least 1000 seconds.
4.4.1 Varying mobility and the number of connections
The purpose of the first set of experiments is to study the impact of host mobility. The
pause time is varied over the range of{0, 50, 100, 200, 300, 500} seconds. Zero pause
time results in the highest mobility since hosts keep moving without a pause. For these
experiments, 10, 20, and 30 connections, which represent light, moderate, and heavy com-
munication requests respectively2, are used. The packet sending rate for each connection
is 4 packets/s. The results are shown in Figure 4.1.
2Traffic load is represented by the sending rate in this chapter. It has different effect on packet loss comparedwith communication request.
38
0 100 200 300 400 5000
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5x 10
6
Pause Time (s)
Byt
e Lo
ss
Total LossCongestion−related LossMobility−related Loss
(a)
0 100 200 300 400 5000
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5x 10
6
Pause Time (s)
Byt
e Lo
ss
(b)
0 100 200 300 400 5000
0.5
1
1.5
2
2.5x 10
7
Pause Time (s)
Byt
e Lo
ss
(c)
0 100 200 300 400 5000
0.5
1
1.5
2
2.5x 10
7
Pause Time (s)
Byt
e Lo
ss
(d)
0 100 200 300 400 5000
0.5
1
1.5
2
2.5
3
3.5
4
4.5x 10
7
Pause Time (s)
Byt
e Lo
ss
(e)
0 100 200 300 400 5000
0.5
1
1.5
2
2.5
3
3.5
4
4.5x 10
7
Pause Time (s)
Byt
e Lo
ss
(f)
Figure 4.1. Packet loss for 4 packets/s CBR connections
39
Packet Loss for AODV
The total packet loss grows from about 3000 to 8000 with the increase of the pause
time from 0 to 500 seconds for 10 connections, as shown in Figure 4.1a. In the case
there are 20 connections, the total packet loss gradually increases by 10% (Figure 4.1c).
For 30 connections, it gradually decreases by 10% (Figure 4.1e). As the communication
request grows from 10 to 20, the total packet loss increases by 9 times when the pause
time is 0 seconds, and by 3 times when the pause time is 500 seconds. The increase of the
communication request from 20 to 30 results in doubled total packet loss.
There is almost no congestion-related packet loss when the communication request
is 10. In the other two cases, the packet loss gradually decreases by about a half as the
pause time increases from 0 to 500 seconds. From 10 to 20 and 30 connections, with no
pause time, the packet loss increases to 5000 and 20000 respectively. The percentage with
respect to the total loss increases as well, to 20% and 30% respectively.
Mobility is always the dominant cause for the packet loss. However, the majority
decreases as the communication request increases. When the pause time is 0 seconds, the
percentage of mobility-related loss decreases from about 100% to 70% and 60%, for 10,
20, and 30 connections. The absolute value and the percentage of the mobility-related
packet loss increase with the pause time.
Packet Loss for DSDV
The growth of the total packet loss with the pause time for DSDV follows a similar
pattern as that for AODV. For 10 connections, the total packet loss increases from about
3000 to 10000 as the pause time increases from 0 to 500 seconds (Figure 4.1b). It is nearly
unchanged with the pause time for 20 and 30 connections as shown in Figure 4.1d and 4.1f
(gradually increases by 5% for 20 connections and decreases by 5% for 30 connections).
Increasing the communication request from 10 to 20 makes the total packet loss grow 10
times and 4 times for 0 and 500 seconds pause time, respectively. The increase of the
communication request from 20 to 30, however, only doubles the total packet loss.
40
The percentage of the congestion-related packet loss increases with the communica-
tion request. Congestion begins to be the dominant cause for the packet loss after the
communication request reaches 20 (it results in approximate 50% and 60% of the total
packet loss with 20 and 30 connections, respectively). The loss is fairly stable with the
pause time, but jitters exist.
The mobility-related packet loss increases with the communication request, but slower
than the congestion-related packet loss.
Comparison between AODV and DSDV
The comparison of different packet losses for AODV and DSDV is as follows.
• Total packet loss:The total packet loss for DSDV is always 10% to 20% higher
than that of AODV, regardless the pause time or the number of connections. For
the moderate and heavy communication requests, the total packet loss for DSDV is
more stable than that of AODV with the increase of the pause time.
• Congestion-related packet loss:DSDV loses more packets due to congestion than
AODV. The gap of the congestion-related packet loss between DSDV and AODV
decreases with the growth of the communication request.
• Mobility-related packet loss:AODV has more mobility-related packet loss than
DSDV.
4.4.2 Varying traffic load and traffic type
The second set of experiments illustrate the effect of traffic load and traffic type. The
pause time ranges over{0, 50, 100, 200, 300, 500} seconds. 10, 20 and 30 connections
are used. Both unresponsive traffic and responsive traffic are studied. The packet rate for
CBR connections is 8 packets/s, which injects a reasonable heavy load to the network. We
use the same mobility scenarios and connection configurations for this set of experiments
as for the previous set of experiments to compare the results with the previous ones.
41
CBR connections with 8 packets/s
As shown in Figure 4.2, each curve that represents a different type of packet loss has
similar shape compared with the corresponding one in the previous experiments, but flatter
(i.e., increase and decrease are more gradual).
For AODV, mobility is still the major cause for the packet loss. Congestion plays a
more important role compared with CBR connection with 4 packets/s rate. Increasing the
number of connections has less effect in this set of experiments than in the previous one.
From 10 connections to 20 connections, the total packet loss increases by only about 3
times. From 20 to 30 connections, the increase is less than 2 times. Comparing Figure
4.2a with Figure 4.1a, the total loss increases by 660% for 0 second pause time, and by
200% for 500 seconds pause time. For the moderate and heavy communication requests,
the total packet loss is only tripled or doubled as the packet rate increases from 4 to 8
packets/s.
For DSDV, congestion dominates the packet loss even when there are only 10 con-
nections. The total packet loss increases the same amount as that for AODV when the
communication request increases, with respect to percentage. The total packet loss with
traffic load is almost the same for DSDV and AODV.
For both AODV and DSDV, increasing the communication request has similar impact
on the total packet loss (i.e., more losses) as increasing the traffic load. The increase of
either parameters will result in decreasing the impact of the other parameter. Heavier
communication request or traffic load introduces more congestion-related packet loss.
TCP connections
The number of bytes (the total size of all lost packets), instead of the number of pack-
ets, is used in the experiments with TCP connections. Because both the application data
and the ACK packets, which have different sizes, are treated as data packets by the routing
protocol, the number of bytes is more comprehensive than the number of packets.
42
0 100 200 300 400 5000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2x 10
7
Pause Time (s)
Byt
e Lo
ss
Total LossCongestion−related LossMobility−related Loss
(a)
0 100 200 300 400 5000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2x 10
7
Pause Time (s)
Byt
e Lo
ss
(b)
0 100 200 300 400 5000
1
2
3
4
5
6
7x 10
7
Pause Time (s)
Byt
e Lo
ss
(c)
0 100 200 300 400 5000
1
2
3
4
5
6
7x 10
7
Pause Time (s)
Byt
e Lo
ss
(d)
0 100 200 300 400 5000
1
2
3
4
5
6
7
8
9
10
11x 10
7
Pause Time (s)
Byt
e Lo
ss
(e)
0 100 200 300 400 5000
1
2
3
4
5
6
7
8
9
10
11x 10
7
Pause Time (s)
Byt
e Lo
ss
(f)
Figure 4.2. Packet loss for 8 packets/s CBR connections
43
0 100 200 300 400 5000
2
4
6
8
10
12
14x 10
5
Pause Time (s)
Byt
e Lo
ss
Total LossCongestion−related LossMobility−related Loss
(a)
0 100 200 300 400 5000
2
4
6
8
10
12
14x 10
5
Pause Time (s)
Byt
e Lo
ss
(b)
0 100 200 300 400 5000
2
4
6
8
10
12
14
16
18x 10
5
Pause Time (s)
Byt
e Lo
ss
(c)
0 100 200 300 400 5000
2
4
6
8
10
12
14
16
18x 10
5
Pause Time (s)
Byt
e Lo
ss
(d)
0 100 200 300 400 5000
0.5
1
1.5
2
2.5x 10
6
Pause Time (s)
Byt
e Lo
ss
(e)
0 100 200 300 400 5000
0.5
1
1.5
2
2.5x 10
6
Pause Time (s)
Byt
e Lo
ss
(f)
Figure 4.3. Packet loss for TCP connections
44
Figure 4.3 demonstrates byte loss in TCP connections3. It shows that the congestion-
related loss for both protocols is greatly reduced by the congestion control mechanism.
The total loss decreases with the decrease of mobility. DSDV outperforms AODV in
terms of the total loss. The total loss of DSDV is only half of that of AODV in all test
cases, because the effect of the major cause for DSDV to lose packets (i.e., congestion) is
diminished by the congestion control mechanism.
For AODV, with the decrease of the congestion-related loss, more than 90% of the
total loss is mobility-related. The total effect of mobility and congestion is less than 20%
for DSDV.
To improve throughput, different routing protocols require different mechanisms to
remedy the major causes for packet loss. Specifically, integrating congestion control tech-
niques with DSDV will significantly improve the throughput, as shown in figure 4.3. For
on-demand routing protocols like AODV, fast rediscovery of new routes will reduce the
mobility-related packet loss, and gain higher throughput consequently. S.R. Das et al.
proposed ad hoc on-demand multipath distance vector (AOMDV) protocol to decrease
the route discovery latency [33]. Their results show that AOMDV loses 3-5% less packets
than AODV. T. Goff et al. proposed preemptive routing maintenance algorithms for ad hoc
networks [34]. The proactive route selection and maintenance are added to the on-demand
protocols to reduce the cost in detecting the disconnection and establishing a new route.
4.5 Discussion
Figure 4.1 and 4.2 show that DSDV loses much more packets due to congestion than
AODV. This difference may result from, with a very great chance, the different route main-
tenance schemes used by DSDV and AODV, because both protocols use distance vector to
represent routing information and choose the routes based on the shortest paths. Since the
per connection traffic load is much lighter (less than 8 packets/s, which is 32Kb/s) than
the communication capacity of a host (2Mb/s), the occurrence of congestion indicates that
3The difference of the amounts of bytes sent by AODV and DSDV is smaller than 5%.
45
S
D
S
D
P1 P1
P2 P2
H
Figure 4.4. Shortest path and congestion
connections converge on heavily burdened hosts. The converged traffic load exceeds the
capacity of those hosts. In a mobile ad hoc network, hosts keep moving. The shortest path
between a source and a destination may change as time passes. DSDV requires periodical
updates of routing information. Every host has the most recent knowledge about routes. It
is likely that the path chosen to forward packets is the currently shortest one. In contrast
to DSDV, AODV picks up a path (usually the shortest one) when a host initiates a route
discovery. The host keeps sending packets via this path until it breaks, even if shorter
paths become available after route discovery.
The difference between these two strategies can be illustrated with Figure 4.4, in which
S is a set of source hosts and D is a set of destination hosts. P1 and P2 are two shortest
paths between S and D. Originally, both DSDV and AODV send packets from S to D
through these two paths. At time t, a host H moves in between S and D, and a shorter path
is available. AODV still sends packets via P1 and P2. DSDV, however, sends all packets
through the new path once it finds out the new one is shorter. Congestion may occur at
host H when traffic load exceeds its capacity. This example shows that keeping sending
packets through the shortest path may cause congestion.
4.6 Conclusion
To our knowledge, this work is the first attempt towards a comprehensive investigation
of packet loss in mobile ad hoc networks. The contributions of congestion and mobility
to the total packet loss have been examined. The impacts of host mobility, communica-
46
tion request, traffic load, traffic type, and AODV and DSDV routing protocols have been
studied. The simulation results show:
• Mobility is the dominant cause for AODV, which is responsible for more than 60%
of the total packet loss. For DSDV, more than 50% of the total packet loss is
congestion-related.
• DSDV loses 10% to 20% more packets than AODV does for UDP traffic. For
TCP traffic, the packet loss for DSDV is a half of that for AODV. DSDV outper-
forms AODV because the congestion control mechanism of TCP greatly reduces
the congestion-related loss.
• Increasing the communication request or traffic load has a stronger impact on the
packet loss in the less stressful situation (i.e., 10 connections at a rate of 4 packets/s).
• Host mobility decreases the packet loss for light communication request and traffic
load. This confirms the argument that mobility increases the capacity of ad hoc
networks [35]. For other cases, the packet loss is rather stable with host mobility.
• Always sending packets via the shortest path may cause congestion at a few heavily
burdened hosts.
Inspired by this work, we are interested in investigating the relationship between short-
est path and congestion. We are working on a loss sensitive routing protocol to support
network layer congestion control for both UDP and TCP traffic. Our ultimate goal is to
build a solid foundation for the research on routing and flow control algorithms for mobile
ad hoc networks.
47
5 SAGA: SELF-ADJUSTING CONGESTION AVOIDANCE ROUTING PROTOCOL
5.1 Introduction
A mobile ad hoc network is a collection of mobile nodes that are deployed as a multi-
hop wireless network without the aid of any preexisting infrastructure or centralized ad-
ministration. The network connectivity and functionality are maintained through cooper-
ations among nodes. Ad hoc networks use wireless links, which have significantly lower
capacity than their hardwired counterparts (e.g., 54Mbps for 802.11g versus 9.952Gbps
for OC192). The real throughput of a wireless link is affected by multiple access, fading,
noise, and interference conditions. It is usually lower than the maximum transmission rate.
The aggregated traffic demand easily reaches or exceeds the link capacity. Congestion is
typically the norm rather than exception in ad hoc networks [16].
Current research efforts that address the congestion avoidance/control problem are
based on the principle of conservation of packets [36]. Examples include TCP and its vari-
eties [37–39]. The conventional TCP-type mechanisms use packet loss to infer congestion
and provide per-connection congestion control. In ad hoc networks, wireless transmis-
sion loss (high bit-error rate) and route reconstruction (network partition) are significant
causes for packet loss. They degrade the effectiveness of congestion inference mecha-
nisms [30, 40]. Mechanisms have been proposed to improve TCP’s performance over
wireless and ad hoc networks [41–46]. The essence of TCP congestion control algorithms
is to reduce the sending rate of traffic upon the occurrence of a congestion.
Two characteristics of ad hoc networks are the existence of multiple routes and the
node-based routing. Routing protocols can make use of them to reduce network con-
gestion with little sacrifice in the sending rate of traffic. Routing with load balancing has
been investigated in [47–49]. The idea is to provide extra information, such as a secondary
metric based on the current load on each node, to help distribute traffic load. It prevents
48
a single node from being overwhelmed. In an ad hoc network, the wireless channel is
shared by multiple nodes. They contend for the channel not only for sending but also for
receiving packets because of the hidden terminal problem [50]. The experimental study
in [51] shows that contention for the channel is the primary reason for network conges-
tion. The impact of the channel contention should be taken into account in the congestion
reduction. For example, if the contention is already intense among a node’s neighbors, it
should not be chosen to forward packets even if there is no load on the node itself.
The main thrust of our work is to reduce network congestion by minimizing channel
contentions. The objective is to avoid thehotspots where multiple nodes are in contention
with each other. The global coupling effect of wireless channel access in ad hoc networks
poses a challenge in determining the contentions locally. In addition, traffic load on a node
must be taken into account, as the store-and-forward process may also cause congestion
when the capacity of a node is exceeded. The shorter routes should be given higher priority
because they are less likely to be involved in contentions with other nodes.
Our approach for reducing contention is as follows: (1) A single server queueing sys-
tem is used to model nodes. The impact of channel contention is quantified using the
service time (the time to successfully transmit a packet over the channel). The routing
cost at each node is computed as the estimated delay. It reflects the effects of channel
contention, current load, and expected load in the immediate future. (2) When a node has
recent traffic, statistical methods are used to evaluate the mean of the delay. When no
recent traffic exists, the underlying MAC protocol is analyzed and probability methods are
applied to compute the expectation of delay. (3) The intermediate delay (IMD) routing
metric is proposed to measure the communication delay introduced by the nodes connect-
ing the source and destination. The route with the least intermediate delay will likely
be involved in the least channel contention. (4) The self-adjusting congestion avoidance
(SAGA) routing protocol is designed to reduce network congestion. Lazy route query op-
eration that is presented in Section 5.4 is used by SAGA to accelerate the establishment of
needed routes. Experimental studies are conducted to evaluate the performance of SAGA
and compare it with AODV [2], DSR [3], and DSDV [1] protocols.
49
This research is conducted in the framework of CSMA/CA (carrier sense multiple
access with collision avoidance) paradigm, which is adopted by the widely used IEEE
802.11 standard [52]. For unicast packets, CSMA/CA requires the sender and receiver to
exchange the request-to-send/clear-to-send (RTS/CTS) frames prior to the transmission of
the actual data frame. Broadcast packets are sent out without RTS/CTS. In this chapter,
packets refer to unicast packets unless otherwise stated.
The rest of this chapter is organized as follows. Section 5.2 introduces contention-
based access to shared media, channel spatial reuse, and the idea of ad hoc routing based
on intermediate delay to reduce congestion. Two methods are presented in Section 5.3 to
estimate delay locally. Section 5.4 presents the detail of SAGA protocol. In Section 5.5,
the performance of the proposed protocol is evaluated and compared with AODV, DSR,
and DSDV. The related work is discussed in Section 5.6. Section 5.7 gives analysis and
guidelines resulting from this research.
5.2 Contention-based media access and congestion avoidance
5.2.1 Characteristics of contention-based access to wireless channels
When transmitting a packet through a wireless channel, the nodes within the transmis-
sion range of the sender, called neighbors, will receive it. If a neighboring node is sending
or receiving a packet simultaneously, a collision occurs. The open channels and the use of
CSMA/CA make the contention in ad hoc networks different from that in wired networks.
Figure 5.1 illustrates the difference using a six-node network. A line between two nodes
denotes that they are neighbors. In ad hoc networks, they are within each other’s trans-
mission range; in wired networks, they are attached to the same physical link. Figure 5.1
shows three transmissions T1, T2, and T3. In the wired network, T1, T2, and T3 can start
simultaneously without collision. In the ad hoc network, T2 will contend with T1 because
the receivers B and D are neighbors. At any time, only one transmission is allowed to use
the channel shared by B and D. T1 and T3 can start concurrently as they are not contending
with each other. The locality of contentions enableschannel spatial reuse[53], i.e. the
50
E
D
B
C
A T1
T2
T3F
Figure 5.1. Network topology and flows
same channel in terms of frequency can be used by multiple transmissions at the same
time.
Channel spatial reuse and the multi-hop routing provide a way to reduce contentions.
For instance, if C wants to establish a connection session with F, selecting the route
C→E→F instead of C→D→F will avoid contention between nodes B and D.
5.2.2 Ad hoc routing based on intermediate delay
The following examples illustrate the use of the intermediate delay in ad hoc routing.
For the purpose of demonstration, the following simplification is used to compute the
delay.
• If the capacity of the wireless channel isC, the size of a packet isP , the delay for
sending a packet isP/C. The MAC layer control messages and the queueing delays
are ignored.
• If n nodes are in contention for a channel, each node getsC/n share of the channel
capacity. The delay for sending a packet isnP/C.
More precise estimates are proposed for SAGA protocol in section 5.3.
51
A
B
C
D E
FG
H I
J
2P/C2P/C
P/CP/CP/C
(a)
A
B
C
D E
FG
H I
J
(b)
Figure 5.2. Select a route with presence of other connections
Figure 5.2, 5.3, and 5.4 illustrate route selection, adaption to traffic changes, and adap-
tion to network topology changes. In each figure, we use a ten-node ad hoc network. The
line with an arrow head represents a connection session. In these examples, a connection
between nodes F and G is to be established.
Figure 5.2 illustrates the route selection process in the presence of other connection
sessions. As shown in figure 5.2a, there is an active connection session between A and C
when F wants to establish a connection with G. D is aware of the contention with A and
computes the delay to be2P/C. Similarly, E’s delay is2P/C. The delay computed by
nodes H, I, and J isP/C. The IMD of the route F→D→E→G is 4P/C, while that of the
route F→H→I→J→G is3P/C. The later route is chosen even though it is one hop longer
(Figure 5.2b). This route is better in terms of channel reuse and congestion avoidance. It
introduces a lower end-to-end delay.
Figure 5.3 illustrates the adaption to changes in traffic. At the beginning, there is
no traffic in the network. Every node can make full use of the channel and introduce
a communication delay ofP/C. The shortest route in terms of hop count is chosen to
establish the connection (figure 5.3a), since it introduces the least intermediate delay. After
the establishment of the connection, a new connection session from A to C is established.
This connection will follow its best route A→B→C. The new connection causes channel
52
A
B
C
D E
FG
H I
J
(a)
A
B
C
D E
FG
H I
J
2P/C2P/C
P/CP/CP/C
(b)
A
B
C
D E
FG
H I
J
(c)
Figure 5.3. Adapt to changes in traffic
A
B
C
D E
FG
H I
J
(a)
A
B
C
D E
FG
H I
J
P/CP/C
P/CP/CP/C
(b)
A
B
C
D E
FG
H I
J
(c)
Figure 5.4. Adapt to changes in network topology
contention between A and D as well as C and E. The new delays at D and E are2P/C.
The IMD of the route F→D→E→G is4P/C. The delay at nodes H, I, and J is stillP/C.
The route F→H→I→J→G has become a better choice, as shown in Figure 5.3b. Node F
re-establishes the connection via the new route. Figure 5.3c shows the result after adapting
to the new connection session.
Figure 5.4 illustrates the adaption to changes in network topology. The first two steps
are same as in the example of Figure 5.2. The route F→H→I→J→G is chosen to avoid
congestion (Figure 5.4a). Suppose nodes A and C have moved and are no longer contend-
53
ing with D and E. F will observe that the route F→D→E→G has become better since its
IMD is 2P/C. The connection is re-established as shown in Figure 5.4c.
These examples demonstrate the essential idea of congestion avoidance by using IMD.
For the design of a practical routing protocol, we must consider the following: (1) At the
time a node computes the delay, it may not know the number of neighbors who are con-
tending with it. (2) Due to the locality of contention, access to a wireless channel creates
global coupling effects in the entire network [53]. Even if the number of contending
nodes is known, the share of capacity cannot be predetermined. (3) The successive links
of a route may interfere with each other.
5.3 Delay estimation
Estimating the delay for sending a packet is critical in SAGA protocol. It is impractical
to compute the accurate delay due to the dynamics and complexity of the network. Fur-
thermore, an accurate value is not required because the delay is transient. The proposed
methods calculate an approximation of the delay.
5.3.1 The model
A node can be modeled as a single server queueing system [54]. The following as-
sumptions are made for delay estimation.
• The incoming traffic is localized with respect to time, i.e., in a short period of time,
it obeys approximately the same distribution.
• The channel access is localized with respect to both time and location. If a node
finds that the channel is busy, so do its neighbors.
• A node has a queue of sufficient size.
• The incoming traffic and outgoing traffic are Poisson processes.
54
• The incoming traffic rate and outgoing traffic rate are independent. This assumption
is reasonable because (a) the complexity of channel contentions washes out the de-
pendency and (b) the incoming traffic includes packets coming from other mobile
nodes as well as from upper layer applications.
The assumptions reduce the complexity of the computation, yet result in a reasonably
good estimate of the real delay. The simulation results presented in Section 5.5 show that
SAGA protocol significantly improves the performance of routing by using the proposed
delay estimation methods.
The following notations are used to describe the parameters of the queueing system.
λ: The arriving rate of packets. It is estimated usingNA
∆t. NA is the number of packets
arrived within the time interval∆t.
µ: The service rate, i.e., the number of packets transmitted over the wireless channel
per second. The capacity of the channel and the contention with neighbors deter-
mine this parameter.
TQ: The wait in the queue before a packet is transmitted.
TS: The average service time for transmitting a packet (TS = 1
µ).
TD: The total delay at a mobile node (TD = TQ + TS).
L: The current length of the queue.
If λ ≥ µ, the maximum allowed value of the delay is assigned toTD since the wait
in queueTQ may be arbitrarily large [54]. Otherwise,TQ can be evaluated using equation
5.1 by applying the Little’s law [54]. The equation holds for general distributions ofλ and
µ.
TQ =λ
µ(µ − λ)+ TSL (5.1)
The delayTD is calculated as follows.
TD = TQ + TS
=TS(L + 1) − NA
∆t(TS)2L
1 − NA
∆tTS
(5.2)
55
L, NA, and∆t can be easily computed. Two cases are considered in the estimation of
the service timeTS: a node with recent traffic (i.e., it recently sends out unicast packets
over the wireless channel) and a node without recent traffic.
5.3.2 Node with recent traffic
If a node has transmitted packets recently, the mean value of the service time can be
obtained using the statistical method. LetNS be the number of packets andTB be the time
that the node spent on transmitting packets.TB is less than or equal to∆t because the
node may not be transmitting packets all the time.
TS =TB
NS
(5.3)
The delayTD is computed using equations 5.2 and 5.3 as follows.
TD =(L + 1) TB
NS− NA
∆tL( TB
NS)2
1 − NA
∆tTB
NS
=(L + 1) − LNA
∆tTB
NS
NS
TB− NA
∆t
(5.4)
To estimateTD, we only need to know the number of incoming and outgoing packets,
current queue length, and the time during which the node is sending packets.
5.3.3 Node without recent traffic
No recent traffic on a node does not imply that a packet can be sent out with the
smallest delay. This is because the neighbors may be using the channel. The expectation
of the service time can be determined by using probability methods to study the procedure
of packet transmission. We analyze the IEEE 802.11 distributed coordination function
(DCF) [52]. Such determination is applicable to other MAC protocols.
Figure 5.5 illustrates the procedure of transmitting a unicast packet using RTS/CTS.
The corresponding state transition is shown in Figure 5.6. We briefly review it for the
purpose of evaluating the expectation of transmission time. The detailed description of
the process is available in [52].
56
����NAV(RTS)
���NAV(CTS)
Sender
Receiver
Others
DIFS Backoff -WindowRTS
CTS
DATA
ACK
Defer access
SIFS
SIFS
SIFS
Figure 5.5. Transmission of a unicast packet using RTS/CTS in the IEEE802.11 standard
When a packet is ready to transmit, the sender picks up a random backoff time of
b × Tslot after observing an idle channel for the time periodTDIFS. b is a random number
uniformly distributed over[0, CW ]. Tslot andTDIFS are values specified by the physical
layer. The sender starts to transmit the RTS frame when the backoff time reaches zero.
The receiver transmits a CTS frame after timeTSIFS upon receiving the RTS frame, if the
media is idle. The neighbors of the sender and receiver set the network allocation vector
(NAV) correspondingly to indicate that the media is reserved. The sender waits for time
TSIFS after receiving the CTS frame and then transmits the data. The receiver waits for
time TSIFS after receiving the data and replies with an acknowledge (ACK) frame. The
expectation of the transmission time for a successful attempt is
E[Tsucc] = TRTS + TCTS + TDATA + TACK + 3TSIFS + E[Tbackoff ] (5.5)
57
Channel is busy
channel for TDIFS
procedureBackoff
Channel is idle,decrease backoff time
Ready totransmit
Suspension
Success
Channel is busy
Fail
TransmissionSucceed
Backoff time reaches zero
Observe an idle
Figure 5.6. State transition of transmission procedure
whereTDATA, TRTS, TCTS andTACK are, respectively, time for transmitting a data packet,
a RTS frame, a CTS frame, and an ACK frame.TDIFS andTSIFS are DCF interframe and
short interframe time, andTbackoff is the time spent on the backoff procedure.
The attempt fails if a CTS frame has not been received at the end ofTtimeout period
following the transmission of the RTS frame. The sender then restarts this process. The
expected time spent on a failed attempt is
E[Tfail] = TRTS + Ttimeout + E[Tbackoff ] (5.6)
Now we compute the expected time spent on the backoff procedureE[Tbackoff ]. According
to the assumption 2 in section 5.3.1, the probability that a channel is busy in a period of a
unit time (i.e., the smallest time unit in the MAC specification) will not change during the
transmission period. It is denoted asp. Observing an idle channel for timet is a Bernoulli
trial [55]. It stops if the channel has been idle for a continuous timet. Let Tidle(t) be the
58
time needed for this trial. The expectation ofTidle(t) is computed recurrently as follows
for a givenp.
E[Tidle(t)] = E[Tidle(t − 1)] + (1 − p) ∗ 1 + p ∗ (1 + E[Tidle(t)]) (5.7)
Equation 5.7 aggregates two cases, assuming that the channel has been idle for a contin-
uous timet − 1 in the trial. (1) The channel is idle in the next unit of time and the trial
stops. The probability for this case is1−p. (2) Otherwise, the trial still needs timeTidle(t)
to stop. Simplifying equation 5.7 results in
E[Tidle(t)] =E[Tidle(t − 1)] + 1
1 − p
Solving the recurrence formula with the initial conditionE[Tidle(1)] = 1
1−pyields
E[Tidle(t)] =1 − ( 1
1−p)t+1
1 − 1
1−p
=( 1
1−p)t − 1
p+ 1
Let Tbackoff(b) be the time needed for a backoff procedure reaching0 from b × Tslot. It is
Tbackoff(b) = Tidle(TDIFS) +b
∑
i=1
t0, t0 = Tslot or Tslot + Tidle(TDIFS + Tslot)
t0 is the time needed to decrease the backoff time byTslot. There are two cases.
1. During the backoff procedure, if no channel activity is detected for the duration of
a particular backoff slot, the backoff time is decreased byTslot. The probability is
(1 − p)Tslot. In this case,t0 = Tslot.
2. Otherwise, the procedure is suspended without decreasing the backoff time. It
resumes after observing an idle channel for timeTDIFS. To decrease the back-
off time, the channel must be idle for a continuous timeTDIFS + Tslot. t0 =
Tslot + Tidle(TDIFS + Tslot).
The expectation oft0 is
E[t0] = (1 − p)TslotTslot + (1 − (1 − p)Tslot)(Tslot + E[Tidle(TDIFS + Tslot)])
59
For a given random numberb, the expectation of time spent on the backoff procedure is
E[Tbackoff(b)] = E[Tidle(TDIFS)] +b
∑
i=1
E[t0]
= E[Tidle(TDIFS)] + bE[t0]
The expected time for the backoff procedure during the transmission attempt is
E[Tbackoff ] = E[E[Tbackoff(b)]]
= E[Tidle(TDIFS)] +CW
2E[t0] (5.8)
The parametersCW (contention window),aCWmin, andaCWmax are defined in the
IEEE 802.11 standard [52].CW takes an initial value ofaCWmin for the first at-
tempt. Every time an attempt fails,CW takes the next value in the series, until it reaches
aCWmax. A successful attempt resetsCW to aCWmin. TheCW values are powers
of 2 minus 1, sequentially ascending fromaCWmin to aCWmax. They are specific to
the physical layer. For example, direct sequence spread spectrum (DSSS) physical layer
management information base (MIB) setsaCWmin to 31 andaCWmax to 1023. In the
rest of this section, we assume DSSS is used as the physical layer. LetCW n be theCW
of the n-th attempt to transmit, then
CW n =
2n+4 − 1, 1≤ n≤ 6;
210 − 1, n > 6.
Let T nsucc andT n
fail be the time spent on a successful transmission and a failed transmission
for the n-th attempt, respectively. From equations 5.5, 5.6, and 5.8, we have
E[T nsucc] = TDATA + TRTS + TCTS + TACK + 3TSIFS + E[Tidle(TDIFS)] +
CW n
2E[t0]
E[T nfail] = TRTS + Ttimeout + E[Tidle(TDIFS)] +
CW n
2E[t0]
The receiver gets the RTS frame if there is no collision during the transmission of
the frame. It will transmit a CTS frame after timeTSIFS if the NAV indicates that the
channel is idle. Otherwise, the receiver will not respond to the RTS frame. The channel
must be idle in this duration ofTRTS + TSIFS for a successful RTS/CTS exchange. Since
60
channel access has locality characteristic, the possibility of a successPs is approximately
(1 − p)TRTS+TSIFS .
The expected transmission time makes sense only for successfully delivered data pack-
ets. We assume that there is no limit on retry and the sender will keep trying until the
packet is delivered. The expected transmission time is
E[Ttrans] = PsE[T 1succ] +
∞∑
i=1
((1 − Ps)iPs(E[T i+1
succ] +i
∑
j=1
E[T jfail])) (5.9)
CW n is fixed whenn ≥ 6, so areE[T nsucc] andE[T n
fail]. Solving equation 5.9 yields
E[Ttrans] = E[T 1succ] +
5∑
i=1
((1 − Ps)i(E[T i+1
succ] − E[T isucc] + E[T i
fail]))
+∞∑
i=6
((1 − Ps)iE[T 6
fail])
= E[T 1succ] +
5∑
i=1
((1 − Ps)i(E[T i+1
succ] − E[T isucc] + E[T i
fail]))
+(1 − Ps)6 1
Ps
E[T 6fail] (5.10)
Once the physical layer parameters are determined, the delay of transmitting a packet
TS is characterized byE[Ttrans], which can be estimated by using equation 5.10 with the
given possibility that the channel is busy. The total delay is calculated by applying the
value ofTS to equation 5.2. This computation is done in constant time.
The IMD metric of a route is obtained by aggregating the delays from each intermedi-
ate nodes along the route. In the proposed delay estimations, the impact of active traffic
in the neighborhood is reflected by the service time or the probability of a busy channel.
The estimation of delay can be done without exchanging information with neighbors.
5.3.4 Accuracy of delay estimation
To evaluate the accuracy of the proposed delay estimation methods, two nodes A and
B are put in an ad hoc network that has active traffic. Node A randomly sends dummy
packets, which only have the delay information, to node B. The rate is 4 packets per 10
seconds, so that the evaluation has little impact on the real traffic. The delay for sending
61
100 200 300 400 500 600 700 800 9000
1
2
3
4
5
6
7
8
9
10
Simulation Time (s)
Del
ay (
ms)
Mean of Estimated Delay = 3.38 msMean of Measured Delay = 3.57 ms
Measured Delay for One PacketMean of Measured Delay in The IntervalEstimated Delay in The Interval
(a) Sender without recent traffic
100 200 300 400 500 600 700 800 900 10000
1
2
3
4
5
6
7
8
9
10
Simulation Time (s)
Del
ay (
ms)
Mean of Estimated Delay = 3.67 msMean of Measured Delay = 3.70 ms
Measured Delay for One PacketMean of Measured Delay in The IntervalEstimated Delay in The Interval
(b) Sender with recent traffic
100 200 300 400 500 600 700 800 9000
1
2
3
4
5
6
7
8
9
10
Simulation Time (s)
Del
ay (
ms)
Mean of Estimated Delay = 3.44 msMean of Measured Delay = 3.73 ms
Measured Delay for One PacketMean of Measured Delay in The IntervalEstimated Delay in The Interval
(c) Sender without recent traffic (receiver is send-
ing packets to other nodes)
100 200 300 400 500 600 700 800 9000
1
2
3
4
5
6
7
8
9
10
Simulation Time (s)
Del
ay (
ms)
Mean of Estimated Delay = 3.74 msMean of Measured Delay = 3.74 ms
Measured Delay for One PacketMean of Measured Delay in The IntervalEstimated Delay in The Interval
(d) Sender with recent traffic (receiver is sending
packets to other nodes)
100 200 300 400 500 600 700 800 9000
1
2
3
4
5
6
7
8
9
10
Simulation Time (s)
Del
ay (
ms)
Mean of Estimated Delay = 3.34 msMean of Measured Delay = 3.65 ms
Measured Delay for One PacketMean of Measured Delay in The IntervalEstimated Delay in The Interval
(e) Sender without recent traffic (sender is receiv-
ing packets from other nodes)
100 200 300 400 500 600 700 800 9000
1
2
3
4
5
6
7
8
9
10
Simulation Time (s)
Del
ay (
ms)
Mean of Estimated Delay = 3.70 msMean of Measured Delay = 3.76 ms
Measured Delay for One PacketMean of Measured Delay in The IntervalEstimated Delay in The Interval
(f) Sender with recent traffic (sender is receiving
packets from other nodes)
Figure 5.7. Comparison of estimated delay and measured delay
62
a dummy packet is not used in the statistics based delay estimation. In the experiments
with recent traffic, data packets are generated from node A to node B besides the dummy
packets.
Experiments have been conducted in six situations. The results are shown in Figure
5.7. It can be observed that for the statistics based delay estimation, the relative error
between the estimated value and the measured value is less than 1.5%. The relative error
for the probability based estimation is less than 8.5%. Both methods produce an estimate
that is smaller than the real delay because of the assumptions we made about the packet
arrival and sending processes.
5.4 Self-adjusting congestion avoidance routing protocol
5.4.1 Introduction
The self-adjusting congestion avoidance (SAGA) routing protocol is designed based
on the ideas presented in sections 5.2 and 5.3. SAGA is a distance vector routing proto-
col. One of the major differences between SAGA and other distance vector based routing
protocols is that SAGA uses IMD instead of hop count as the distance. It gives SAGA the
capability of balancing traffic load and dealing with congestion.
To send packets, every node maintains a routing table that contains entries to all known
nodes in the network. The data structure of the routing entry is shown in Figure 5.8.seq is
a field of a routing entry that stores the sequence number representing the “freshness” of a
route as in DSDV and AODV. It is maintained by the destination. Routes with more recent
sequence numbers are always preferred for routing decisions. For routes with identical
sequence number, the one with the smallest IMD is chosen.
SAGA is a proactive protocol like DSDV, which requires every node to periodically
advertise the routing table to its neighbors. Significant new information such as a new
route or a broken route may also trigger an advertisement. The estimated delay at this
node is included in advertisement packets. A broken or unavailable route is assigned a
63
class RTEntry {RTEntry();addr_t dst; // destinationaddr_t next_hop; // next hopfloat imd; // routing metric (IMD)uint seq; // sequence number// minimum value of imd in all advertisements// since the last update of seqfloat min_advertised_imd;// when is ok to advertise this route?time advertise_ok_at;// do we need to advertise this route?bool need_advertise;// number of MAC callbacksuint MAC_callback_cnt;// event indicating this route breaksEvent* trigger_event;// event indicating the next hop is not availableEvent* timeout_event;PacketQueue* q; // packet queue
};
Figure 5.8. Data structure of the routing entry
Table 5.1Major constants of SAGA protocol
Variable Meaning Value
MIN INTERVAL minimum time between two advertisements1 seconds
MAC CALLBACK how many callbacks indicate a broken link 2
STARTUPADVERTISE advertisements sent during startup 5
PERIODADVERTISE time between two full advertisements 15 seconds
DEFAULT TTL default value of TTL 30
delay of∞, which is a value greater than the maximum allowed value of the delay. A
route with∞ delay is considered as invalid and is usually not included in advertisements.
Table 5.1 shows the major constants of SAGA and their values that are used in the
simulation study.
64
Delay EstimatingTimer
MAC SamplingTimer
MAC Callback
With activetraffic?
Compute delay using
NA, t, NS, TB, and L
Computer delay using
NA, t, Nsample, Nbusy,and L
Store estimated delayReset local variables
EstimatedDelay
Nsample++
Channelbusy?
Nbusy++
A packetarrives?
Begin to senda packet?
End of sendinga packet?
NA++
starttime = now
NS++TB += now-starttime
L = MAC layer queue length
set timer to t
YES
NO
YES
NO
YES
NO
YES
NO
YES
r = random numberbetween 1 and 500
clear timer
set timer to r*Tslot
Figure 5.9. Delay estimate
5.4.2 Operations
SAGA protocol uses the following operations to estimate the delay, advertise and
maintain routes, and handle broken links. The packet forwarding procedure is similar
to that in DSDV. It is not discussed in this chapter.
Delay estimation
Every node estimates the delay using the methods presented in Section 5.3. The num-
ber of arrived packetsNA, the number of sent packetsNB, the time during which the node
is transmitting packetsTB, the current length of the queueL, and the probability of a busy
channelp are needed for estimation. The duration∆t determines how frequently the delay
is estimated. It is set to the time interval between two full advertisements. NA, NB, and
TB are counted using a MAC callback function. This function is invoked when a packet
arrives at MAC, when a packet is ready to transmit, and after a packet is transmitted. The
65
probabilityp is determined by using randomly sampling.p =Nbusy
Nsample, whereNsample is
the number of samples andNbusy is the number of samples that indicate a busy channel.
Each node maintains a sampling timer, which will be randomly triggered about 200 times
per second. When the timer is triggered, SAGA checks state of the channel. This timer is
set when a new estimate process begins. It is cleared if active traffic is detected.
Route advertisement
SAGA uses route advertisements to disseminate information throughout the network.
Two types of advertisements are defined in SAGA protocol. One, called “full advertise-
ment”, carries all the available routing information. The other, called “partial advertise-
ment”, carries only information changed since the last advertisement. Full advertisements
are generated relatively infrequently. If a partial advertisement contains most of routing
entries, it is upgraded to a full advertisement so that the next partial advertisement will
be smaller. Two events will trigger an advertisement. The first one triggers a full ad-
vertisement, which is scheduled PERIODADVERTISE seconds after the previous full
advertisement. In the bootstrapping phase, a node may schedule full advertisements more
frequently. The other event triggers a partial advertisement upon receiving significant new
routing information, including: (1) a more recent sequence number, which helps SAGA
to adapt in circumstances similar to the example shown in Figure 5.3; (2) a broken link, as
discussed in [56], propagating bad news quickly will improve system performance.
Each routing entry is associated with two flags:needadvertiseandadvertiseok at.
A partial advertisement only contains those entries whoseneedadvertiseis set. A full
advertisement includes all entries whoseadvertisementok at is earlier than the current
time. In both cases, the estimated delay that the node may introduce and the sequence
number are included in the advertisement packets. Before each full advertisement, the
sequence number is incremented by 2 so that the sequence number maintained by the
destination is always an even integer. The pseudo code for making an advertisement packet
is shown in Figure 5.10.
66
MakeAdv(periodic) {for each routing entry rte
entry_count++;if rte.need_advertise == TRUE
count++;if rte.advertise_ok_at > now
unadvertiseable++;if count >= entry_count*2/3
periodic = TRUE;make an advertisement packet p;add estimated delay to p;if periodic == TRUE
increment sequence number by 2;add sequence number to p;add entry_count - unadvertiseable to p;for each routing entry rte
if rte.advertise_ok_at <= nowadd rte to p;rte.need_advertise = FALSE;
elseadd sequence number to p;add count to p;for each routing entry rte
if rte.need_advertise == TRUEadd rte to p;rte.need_advertise = FALSE;
}
Figure 5.10. Algorithm for making an advertisement packet
Route maintenance
Each routing entry in SAGA has two fields associated with the distance: (a)imd stores
the intermediate delay and is used for route advertisement; and (b)min advertised imd
stores the minimum value ofimd in all advertisements since the last update ofseq. As-
sume that a nodei receives an advertisement of a route to a nodex from a neighborj.
The route has a sequence numberseqxj and an intermediate delayimdx
j . Nodei updates
its routing table if and only if one of the following four conditions is true.
1. Nodei does not have a valid route to the destinationx.
2. Nodej is the next hop of the current route.
3. The new route contains a fresher (valid) sequence number (seqxj > seqx
i ) and
imdxj < ∞.
4. seqxj = seqx
i andmin advertised imdxi > imdx
j .
67
These constraints guarantee that SAGA will not introduce loops in routes. In the third
condition, as proved in [1], a loop cannot be created if nodes use fresher sequence numbers
to pick routes. The loop-free property holds in the fourth condition due to the theorem
proved in [57], which states that distance vector algorithms always maintain loop-free
routes in presence of static or decreasing link weights.
Some distance vector routing protocols use a single field of distance for a routing entry.
This field is used for both route advertisement and routing decision. Becauseimd reflects
the extent of congestion along a route, its value may change even if the route is static. As
imd is not static or decreasing, the loop-free property may not hold if it is used to make
routing decisions as in the fourth condition. The use ofmin advertised imd assures
loop-free routes.
Adaptive routing metrics such asimd sometimes suffer from oscillation. After choos-
ing a route and beginning to send packets, other routes become attractive. The tendency of
the routing decision to switch excessively from one choice to alternates makes the routes
unstable. This increases routing overhead and decreases performance. The oscillation
problem was stated in [58]. Usingmin advertised imd in route decision prevents a node
from switching back and forth among alternative routes and helps in reducing the oscil-
lation of theimd value. The associated cost is a possible delay in adopting a better route
whose intermediate delay is lower thanimd but higher thanmin advertised imd.
Figure 5.11 demonstrates the procedure of route maintenance. An extra functionality
of route maintenance is shown in functionProcessAdvEntry. The reception of an adver-
tisement entry with an older sequence number will trigger a partial advertisement to help
the neighbor to obtain the up-to-date route.
Handling broken links
A link to a neighbor is considered as broken if no advertisement is received from it for
a certain period of time. To detect broken links, each neighbor is associated with a timeout
68
ProcessAdv(pkt) {sender = source address of pkt;delay = estimate delay in pkt;seq = sequence number in pkt;rte = routing entry to sender;if rte does not exist
add a routing entry rte;rte.dst = rte.next_hop = sender;rte.imd = 0;rte.seq = seq;rte.need_advertise = TRUE;trigger advertisement for rte;
elserte.next_hop = sender;rte.imd = 0;if (rte.seq < seq)
rte.seq = seq;rte.need_advertise = TRUE;trigger advertisement for rte;
for each advertisement entry advif adv.dst == my_address
if adv.imd != 0schedule a full advertisement;
elseadv.imd = adv.imd + delay;adv.next_hop = sender;ProcessAdvEntry(adv);
}
ProcessAdvEntry(adv) {rte = routing entry to adv.dst;if rte does not exist
add a routing entry rte;rte.dst = adv.dst;rte.next_hop = adv.next_hop;rte.imd = adv.imd;rte.seq = adv.seq;rte.need_advertise = TRUE;trigger advertisement for rte;
else if rte.seq == adv.seqif rte.min_advertised_imd > adv.imd
UpdateRoute(rte, adv);else if rte.seq < adv.seq
if adv.imd < INFINITY or rte.next_hop == adv.next_hopUpdateRoute(rte, adv);rte.need_advertise = TRUE;trigger advertisement for rte;
else if rte.seq > adv.seqif rte.imd < INFINITY and adv.imd == INFINITY
rte.need_advertise = TRUE;trigger advertisement for rte;
}
Figure 5.11. Algorithm for route maintenance
69
MACCallback(pkt) {next_hop = next hop of pkt;drop pkt for MAC callback;rte = routing entry to next_hop;rte.MAC_callback_cnt++;if rte.MAC_callback_cnt > MAC_CALLBACK
if rte.timeout_event existscancel rte.timeout_event;
HandleTimeout(rte);}
HandleTimeout(rte) {for each routing entry rte2
if rte2.next_hop == rte.dst and rte2.imd < INFINITYrte2.imd = INFINITY;rte2.seq++;rte2.need_advertise = TRUE;trigger advertisement for rte2;
}
Figure 5.12. Algorithm for handling broken links
event that will be triggered after2 × PERIOD ADV ERTISE seconds. This event is
reset every time an advertisement is received from the neighbor.
The MAC callback is another mechanism to detect broken links, because CSMA/CA
will report an error when it fails to transmit a packet. Continuous occurrence of failure
indicates that either the neighbor is not available or the contention is too intense. In both
cases, this neighbor should not be picked as the next hop. If the number of continuous
callbacks exceeds the preset threshold MACCALLBACK, the timeout event is triggered.
Figure 5.12 shows how MAC callback triggers the event and how SAGA handles a
broken link. When a link to a neighbor is indicated broken, any route through that neighbor
is immediately assigned∞ to IMD and the sequence number is incremented by 1. Thus
a broken route is always associated with an odd sequence number while the valid one is
associated with an even sequence number.
Lazy route query
SAGA does not provide a dedicated route query operation as in the on-demand proto-
cols. When a node wants to send packets to a destination but does not have a valid route,
70
it uses a technique calledlazy route query. Usually, a route with∞ delay in an advertise-
ment packet is used to report a broken link. In this case,seq is an odd number. A route
with ∞ delay and an even number ofseq is treated as a query instead of an advertisement.
It indicates that this node needs a route to the destination. The route’sseq must be greater
than the one in the query. Neighbors who have a valid route will include it in the next
advertisement packet as a response to the query. Lazy route query works well with the
proactive approach, because (1) each node periodically advertises its routing table, it is
likely that one of the neighbors has already had a valid route; and (2) multiple routes may
be queried in one advertisement packet.
These operations enable SAGA protocol to handle the dynamic and unpredictable
changes in the network topology and traffic load, and to deliver packets through routes
with less congestion. They are the basis of a complete implementation for experimental
studies. Please refer to [59] for the details of SAGA protocol.
5.5 Experimental evaluation
The objective of the experiments is to study the performance of routing protocols un-
der congestion. SAGA is compared with AODV, DSR, and DSDV protocols, which have
received wide attention in the literature [8,9,25]. The use of intermediate delay in SAGA
is contrasted against the use of hop count in AODV, DSR, and DSDV through the mea-
surements obtained from the experiments.
The study is done through simulation using the network simulator ns2 [18]. The imple-
mentations of AODV, DSR and DSDV protocols are provided by ns2. All optimizations
for AODV and DSR are enabled in the simulation for the comparisons. The implemen-
tation of SAGA is based on the operations presented in Section 5.4. The values of the
parameters for SAGA are given in Table 5.1 (Major constants of SAGA protocol).
The wireless interface simulates the 914 MHz Lucent WaveLAN direct-sequence spread-
spectrum (DSSS) radio interface [7]. The IEEE 802.11 distributed coordination func-
tion (DCF) with CSMA/CA is used as the MAC layer protocol. The random waypoint
71
model [9] is used to generate movements for mobile nodes. These settings are commonly
used in studies reported in the literature.
Five independent scenarios are generated for each experiment. The average values are
used for analysis.
5.5.1 Performance metrics
The following metrics are used to evaluate the routing protocols. They are based on a
list of quantitative metrics suggested by the RFC 2501 [16].
• Delivery Ratio: The ratio of the data delivered to the destinations (i.e., throughput)
to the data sent out by the sources. The throughput is also studied in the experiments.
• Protocol Overhead: The ratio of the routing load to the data successfully delivered
to the destination. The routing load is measured as the number of bytes of protocol
messages transmitted hop-wise. The transmission on each hop is counted once.
• Average End-to-end Delay: The average time it takes for a packet to reach the des-
tination. It includes all possible delays in the source and each intermediate node. It
can be caused by routing discovery, queueing at the interface queue, transmission at
the MAC layer, etc. Only successfully delivered packets are counted.
5.5.2 Simulation and input parameters
UDP connections are used in most of the experiments so that there is no congestion
control at the transport layer. Each connection is specified as a randomly chosen source-
destination (S-D) pair. Every connection starts at a time uniformly distributed over 0 to
100 seconds, so that the proactive protocols have sufficient to warm up. The size of packets
is 512 bytes. Two types of traffic are considered in the study.
• Constant Bit Rate (CBR) traffic:It is generated at a deterministic rate [18]. This
type of traffic is widely used in the study of ad hoc network routing protocols and
provides a good basis for evaluating SAGA protocol.
72
100 120 140 160 180 200 220 240 260 280 3000
50
100
150
200
250
300
Simulation Time (seconds)
Tra
ffic
Load
(kb
/s)
Figure 5.13. POO traffic
• Pareto On/Off (POO) traffic:It is generated according to a pareto on/off distribu-
tion [18]. Packets are sent at a fixed rate during on periods, and no packets are sent
during off periods. Both on and off periods are drawn from a pareto distribution.
POO traffic exhibits long range dependency. It closely matches with the empirically
measured network traffic [60]. Figure 5.13 shows an example of the aggregated
POO traffic used in the simulation.
To highlight the impact of congestion on the routing performance, the offered traffic
load is taken as the input parameter. The aggregated traffic injected into the network
ranges from 100 kb/s to 600 kb/s, which puts much stress on the routing protocols.
Six experiments for UDP connections and two experiments fot TCP connections have
been conducted connections by varying the maximum speed of the movement of nodes
and the number of connections. The maximum speeds of 4m/s and 20m/s are considered
as low and high mobility respectively. The first four experiments use CBR traffic and the
last two use POO traffic.The values of parameters used in the simulation are given in Table
5.2 (Simulation and input parameters).
73
Table 5.2Simulation and input parameters
Simulation time 1000 seconds
Independent runs 5
Mobility model random waypoint
Simulation area 1000m× 1000m
Maximum speed 4 m/s, 20 m/s
Pause time 10 seconds
Wireless transmission range 250m
Channel capacity 2 Mb/s
Number of mobile nodes 50
Number of connections 10, 30
Packet size 512 bytes
POO on time 500 ms
POO off time 1000 ms
POO shape 1.5
5.5.3 Measurements and observations
Experiment 1
The experiment studies the routing performance in low mobility environment. Twenty
percent of the nodes are generating CBR traffic. Data in Figure 5.14 shows that the per-
formance of the routing protocols decreases with the increase of the offered traffic load.
Figure 5.14b shows that the increase in traffic has more impact on AODV and DSR
than on SAGA and DSDV. When the offered traffic load increases from 100 kb/s to 500
kb/s, the delivery ratios of AODV and DSDV drop from 96% to 56% and from 92% to
47% respectively. The delivery ratios of SAGA and DSDV stay stable when the offered
traffic load is less than 300 kb/s. SAGA delivers around 95% of the packets, while DSDV
74
0 100 200 300 400 5000
50
100
150
200
250
300
350
400
450
Offered Traffic Load (kb/s)
Thr
ough
put (
kb/s
)
AODVDSDVDSRSAGA
(a)
0 100 200 300 400 5000
20
40
60
80
100
Offered Traffic Load (kb/s)
Del
iver
y R
atio
(%
)
(b)
0 100 200 300 400 5000
50
100
150
200
Offered Traffic Load (kb/s)
Pro
toco
l Ove
rhea
d (%
)
(c)
0 100 200 300 400 5000
0.5
1
1.5
2
Offered Traffic Load (kb/s)
Ave
rage
End
−to
−en
d D
elay
(s)
(d)
Figure 5.14. 10 CBR connections, low mobility
delivers 85%. The delivery ratios drop to 77% and 71% respectively, when the offered
traffic load reaches 500 kb/s.
Figure 5.14c shows that the overhead of AODV and DSR increases with the offered
traffic load, from 13% to 85% and from 14% to 57%, while that of DSDV decreases from
72% to 35%. The overhead of SAGA drops from 32% to 12% when the offered traffic
load increases from 100 kb/s to 300 kb/s. It then increases to 23%.
The average end-to-end delay of DSR increases significantly from 0.15 to 1.16 sec-
onds. The delay of AODV gradually increases from 0.03 to 0.2 seconds. SAGA and
DSDV have almost the same delay of 0.02 seconds when the offered traffic load is less
75
0 100 200 300 400 5000
50
100
150
200
250
300
350
400
450
Offered Traffic Load (kb/s)
Thr
ough
put (
kb/s
)
AODVDSDVDSRSAGA
(a)
0 100 200 300 400 5000
20
40
60
80
100
Offered Traffic Load (kb/s)
Del
iver
y R
atio
(%
)
(b)
0 100 200 300 400 5000
50
100
150
200
Offered Traffic Load (kb/s)
Pro
toco
l Ove
rhea
d (%
)
(c)
0 100 200 300 400 5000
0.5
1
1.5
2
Offered Traffic Load (kb/s)
Ave
rage
End
−to
−en
d D
elay
(s)
(d)
Figure 5.15. 10 CBR connections, high mobility
than 300 kb/s. The delay of DSDV then increases to 0.25 seconds, while that of SAGA
increases to 0.11 seconds (Figure 5.14d).
Experiment 2
In this experiment, all the parameters are the same as in the previous experiment,
except that the maximum moving speed is changed from 4 m/s to 20 m/s.
The delivery ratio of DSR drops quickly with the increase of the offered traffic load
(Figure 5.15b), because its throughput stays at around 100 kb/s as shown in figure 5.15a.
76
The overhead of DSR increases sharply compared with results of the low mobility exper-
iment, by 2 to 5 times depending on the offered traffic load (Figure 5.15c and 5.14c).
With the offered traffic load increasing from 100 to 500 kb/s, the delivery ratios of
SAGA and DSDV drop from 89% to 74% and from 77% to 65% as shown in Figure 5.15b.
When mobility changes from low to high, the overhead of DSDV is almost doubled, and
that of SAGA slightly increases by 5%.
Comparing Figure 5.15 with figure 5.14, we can tell that mobility greatly affects the
performance of DSR. For SAGA and DSDV, the increase of mobility has a greater im-
pact when the offered traffic load is lighter. Mobility does not have much impact on the
performance of AODV.
Experiments 3 and 4
These two experiments illustrate the performance of routing protocols when the num-
ber of connections is 30. In the simulation, 60% of the mobile nodes are generating traffic.
The aggregated traffic load is the same. The results of the low mobility experiment are
shown in Figure 5.16 and those of the high mobility experiment are shown in figure 5.17.
Comparing Figure 5.16 with figure 5.14, and Figure 5.17 with figure 5.15, we can
conclude that routing performance decreases with the number of connections, which has a
greater impact on AODV and DSR, the on-demand protocols, than on SAGA and DSDV.
In the low mobility experiment, the delivery ratio of each protocol is almost unchanged
with the increase of the number of connection when the offered traffic load is light (less
than 200 kb/s). It decreases by less than 10% with 500 kb/s traffic. Unlike the other three
protocols, the overhead of AODV increases significantly. The average delay increases for
each protocol, but AODV has the greatest growth.
In the high mobility experiment, the throughput of DSR is saturated at 100 kb/s, almost
the same as in the experiment with 10 connections. The throughput of AODV is saturated
at about 200 kb/s, while the saturation is not obvious in the corresponding 10-connection
experiment (Figure 5.17a and 5.15a).
77
100 200 300 400 500 6000
50
100
150
200
250
300
350
400
450
Offered Traffic Load (kb/s)
Thr
ough
put (
kb/s
)
AODVDSDVDSRSAGA
(a)
100 200 300 400 500 6000
20
40
60
80
100
Offered Traffic Load (kb/s)
Del
iver
y R
atio
(%
)
(b)
100 200 300 400 500 6000
50
100
150
200
Offered Traffic Load (kb/s)
Pro
toco
l Ove
rhea
d (%
)
(c)
100 200 300 400 500 6000
0.5
1
1.5
2
Offered Traffic Load (kb/s)
Ave
rage
End
−to
−en
d D
elay
(s)
(d)
Figure 5.16. 30 CBR connections, low mobility
Experiments 5 and 6
In addition to CBR traffic, experiments have been conducted to study the performance
of routing protocols using POO traffic. The long range dependency of the aggregated
POO traffic closely matches with the actual network traffic. The study provides a better
understanding on the performance when the routing protocols are implemented for ad hoc
networks in practice. The simulation parameters in these experiments are the same as in
the 10-connection experiments, except that every source of a connection generates POO
traffic instead of CBR traffic. As shown in Figure 5.13, although the average traffic load is
about 133 kb/s, the real-time load often approaches or exceeds 200 kb/s. The fluctuation
78
100 200 300 400 500 6000
50
100
150
200
250
300
350
400
450
Offered Traffic Load (kb/s)
Thr
ough
put (
kb/s
)
AODVDSDVDSRSAGA
(a)
100 200 300 400 500 6000
20
40
60
80
100
Offered Traffic Load (kb/s)
Del
iver
y R
atio
(%
)
(b)
100 200 300 400 500 6000
50
100
150
200
Offered Traffic Load (kb/s)
Pro
toco
l Ove
rhea
d (%
)
(c)
100 200 300 400 500 6000
0.5
1
1.5
2
2.5
Offered Traffic Load (kb/s)
Ave
rage
End
−to
−en
d D
elay
(s)
(d)
Figure 5.17. 30 CBR connections, high mobility
of traffic load poses a challenge that requires a quick response to traffic dynamics. The
results are shown in Figure 5.18 and figure 5.19.
The performance of DSDV and DSR is almost the same as in the 10-connection ex-
periments. SAGA performs even better when the offered traffic load is in the range of 100
kb/s to 300 kb/s. In terms of delivery ratio, SAGA outperforms all evaluated protocols in
all cases except for DSR in the 67 kb/s traffic and high mobility (Figure 5.18b and 5.19b).
AODV delivers less than 40% of packets in the low mobility experiment and about
20% of packets in the high mobility experiment. Figure 5.18c and 5.19c show that the
overhead of AODV is less than 10%, which is much lower compared with the results of
79
0 100 200 300 400 500 6000
50
100
150
200
250
300
350
400
450
Offered Traffic Load (kb/s)
Thr
ough
put (
kb/s
)
AODVDSDVDSRSAGA
(a)
0 100 200 300 400 500 6000
20
40
60
80
100
Offered Traffic Load (kb/s)
Del
iver
y R
atio
(%
)
(b)
0 100 200 300 400 500 6000
50
100
150
200
Offered Traffic Load (kb/s)
Pro
toco
l Ove
rhea
d (%
)
(c)
0 100 200 300 400 500 6000
0.5
1
1.5
2
Offered Traffic Load (kb/s)
Ave
rage
End
−to
−en
d D
elay
(s)
(d)
Figure 5.18. POO traffic, low mobility
the CBR traffic experiments. It indicates that AODV does not exchange much routing
information when traffic bursts. Many packets are dropped due to congestion.
Experiments for TCP traffic
Two experiments have been conducted to evaluate the performance of SAGA with
TCP traffic in low and high mobility scenarios. The results are shown in Figure 5.20.
All evaluated routing protocols except for DSR have almost the same end-to-end delay
regardless the number of connections as shown in Figure 5.20c and 5.20d. The proactive
protocols achieve higher throughput than the on-demand ones. This is consistent with the
80
0 100 200 300 400 500 6000
50
100
150
200
250
300
350
400
450
Offered Traffic Load (kb/s)
Thr
ough
put (
kb/s
)
AODVDSDVDSRSAGA
(a)
0 100 200 300 400 500 6000
20
40
60
80
100
Offered Traffic Load (kb/s)
Del
iver
y R
atio
(%
)
(b)
0 100 200 300 400 500 6000
50
100
150
200
250
Offered Traffic Load (kb/s)
Pro
toco
l Ove
rhea
d (%
)
(c)
0 100 200 300 400 500 6000
0.5
1
1.5
2
Offered Traffic Load (kb/s)
Ave
rage
End
−to
−en
d D
elay
(s)
(d)
Figure 5.19. POO traffic, high mobility
results obtained from the study of packet loss. SAGA still performs better than DSDV,
but not much. This is because TCP also tries to control congestion, thus diminishes the
advantage of SAGA in terms of congestion avoidance.
5.5.4 Analysis and discussion
We classify the traffic load offered by CBR connections into low, moderate, and high
based on whether it is less than 200 kb/s, between 200 and 400 kb/s, or greater than 400
kb/s. For the traffic load offered by POO connections, the two classifying values are 132
kb/s and 330 kb/s.
81
5 10 15 20 2540
60
80
100
120
140
160
180
Number of TCP Connections
Thr
ough
put (
kb/s
)
AODVDSDVDSRSAGA
(a) Maximum speed = 4m/s
5 10 15 20 2540
60
80
100
120
140
160
180
Number of TCP Connections
Thr
ough
put (
kb/s
)
(b) Maximum speed = 20m/s
5 10 15 20 250
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Number of TCP Connections
Ave
rage
End
−to
−en
d D
elay
(s)
(c) Maximum speed = 4m/s
5 10 15 20 250
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Number of TCP Connections
Ave
rage
End
−to
−en
d D
elay
(s)
(d) Maximum speed = 20m/s
Figure 5.20. TCP traffic
SAGA versus on-demand protocols
• Throughput: SAGA is able to sustain heavier traffic load and offers higher peak
throughput than AODV and DSR. Since SAGA can balance traffic load and avoid
congestion, it enables better utilization of the aggregated network capacity. SAGA
provides a peak throughput of 400 kb/s while AODV and DSR saturate at around
250 kb/s, when mobility is low (Figure 5.14a and 5.16a). In high mobility scenarios,
DSR saturates at 120 kb/s. The peak throughput of AODV is 220 to 280 kb/s.
It decreases as the number of connections increases. SAGA can reach about 370
kb/s regardless of the number of connections (Figure 5.15a and 5.17a). POO traffic
82
does not have much impact on the peak throughput of SAGA and DSR. It causes
the peak throughput of AODV to drop to 150 kb/s and 100 kb/s in low and high
mobility scenarios respectively (Figure 5.18a and 5.19a). In summary, SAGA can
consistently offer a peak throughput of 370 to 400 kb/s in all cases, which is 1.5 to
3.5 times of the peak throughput achieved by the on-demand protocols.
• Delivery ratio: SAGA does not achieve high delivery ratio in high mobility and low
traffic load (Figure 5.15b, 5.17b, and 5.19b). More than 95% of the dropped packets
are caused by broken routes, because the routes obtained from advertisements may
be stale by the time they are used. In the implementation of SAGA, a link is con-
sidered broken if two consecutive packets to the same neighbor are dropped. This
increases the accuracy of broken link detection, at the cost of more dropped packets.
When mobility is high and traffic load is low, one packet might be enough to infer
a broken link since the probability of dropping packets due to congestion is low.
In this case, SAGA does not deliver as many packets as AODV and DSR. When
the offered traffic load increases from low to moderate, the delivery ratio of SAGA
increases because of the accuracy of broken link detection.
SAGA performs as well as the on-demand protocols in low traffic load and low
mobility. It outperforms them when the offered traffic load is moderate or high.
• Protocol overhead:The protocol overhead of SAGA is in the range of 15% to 50%
of the total delivered data. Because SAGA uses one-hop broadcast and requires the
interval between two consecutive advertisements to be at least one second, its over-
head is not affected much by traffic load, mobility, and the number of connections.
The overhead of AODV increases rapidly with the offered traffic load. The POO
traffic experiments are exceptions where AODV fails to deliver most of the packets.
AODV uses network-wide broadcast to re-discover a route when a packet is dropped
and the route is considered as broken. With the increase of the offered traffic load,
a larger number of packets are dropped due to congestion. This causes AODV to
initiate additional route re-discoveries. AODV introduces less overhead than SAGA
83
only in low traffic load. In the worst case, the overhead of AODV is as ten times as
that of SAGA (Figure 5.16c).
The overhead of DSR is affected by mobility. It is almost tripled when the maximum
moving speed of nodes changes from 4 m/s to 20 m/s. DSR uses route cache and
snooping that are not effective in highly dynamic networks. Only when mobility
and the offered traffic load are low, DSR can outperform SAGA in terms of protocol
overhead. Otherwise, it introduces up to 8 times overhead as SAGA does (Figure
5.17c).
• End-to-end delay:SAGA offers lower average end-to-end delay than AODV and
DSR, because it uses the intermediate delay instead of the hop count as the routing
metric. The advantage of using the new metric is significant when the offered traffic
load is high. In those cases, the delay of SAGA is 50% less than that of AODV and
80% less than that of DSR (Figure 5.14d, 5.15d, 5.16d, and 5.17d).
SAGA versus DSDV
SAGA outperforms DSDV in the measured metrics at the conducted experiments. It
delivers 10% more packets than DSDV with less than half of the protocol overhead. The
average end-to-end delay of SAGA is almost the same as that of DSDV when the offered
traffic load is less than 300 kb/s. It is around 50% to 70% of the delay of DSDV with 500
kb/s traffic, depending on mobility and the number of connections. DSDV fails to provide
high delivery ratio in low traffic load. It delivers about 85% of the packets while the other
protocols can deliver 95% (Figure 5.14b and 5.16b). In addition, it introduces 1 to 2 times
more overhead than other protocols in high mobility and low traffic load (Figure 5.15c and
5.17c).
Additional experiments have been done with various maximum speeds ranging from 4
m/s to 24 m/s and numbers of connections ranging from 10 to 50. They lead to the similar
conclusions.
84
5.6 Related work
Associativity-based routing (ABR) [47] is one of the first protocols that consider load
as a part of the routing metric. The load is based on the number of routes in which a
node is involved. Load balancing routing protocols [48,49] use a similar idea as ABR but
different methods to compute load. Various traffic loads on different routes have not been
considered.
Multipath routing protocols [33, 61, 62] can be adjusted for load balancing by allow-
ing sources to deliver packets through different paths. The source-based load balancing
may still cause congestion. Even though every single source evenly distributes load over
multiple paths, nodes that are involved in several paths can be overloaded.
A. Boukerche and S.K. Das present a new approach to control congestion in wireless
ad hoc networks in [63, 64]. They propose a randomized version of the DSDV routing
protocol called R-DSDV. R-DSDV propagates the routing messages according to a routing
probability distribution rather than on a periodic basis. It controls congestion in the store-
and-forward procedure. If the current queue size is over the congestion level, a newly
arrived packet is dropped or queued according to a probability. The data packets have
higher priorities than the advertisement packets.
The experimental evidence from two empirical wireless test-beds presented by D.S.J.
Couto, D. Aguayo, B.A. Chambers, and R. Morris in [65] shows that the minimum-hop-
count routing often chooses routes that have significantly less capacity than the best paths
in a multi-hop wireless network. A new metric, the expected transmission count (ETX), is
designed for routing protocols to find high-throughput paths [58]. The expected number
of transmission is determined by the forward and reverse delivery ratios of a link, which
are measured using dedicated link probe packets. The ETX metric incorporates the effect
of link loss ratio and the interference among the successive links of a path. It does not
account for mobility and does not route around congested links. It is complementary to
the IMD metric proposed in this chapter.
85
C. Cordeiro, S.R. Das, and D.P. Agrawal propose contention-based path selection (CO-
PAS) for TCP over multi-hop wireless networks [66]. COPAS monitors the MAC layer
contention and accordingly changes the forward and reverse paths for a TCP connection.
It enhances the performance of TCP by minimizing the likelihood of the capture prob-
lem [67]. The number of backoffs is used to measure contention. Because the number
of backoffs is closely related to the number of packets that are sent during the measured
time, research is needed for a more precise indication of channel contention. Intermediate
nodes continuously piggyback their contention information on packets that pass through
them. If the number of backoffs exceeds a predefined threshold, the route is reconstructed.
In a network with heavy traffic or lossy links that result in a large number of backoffs,
unnecessary route reconstructions can be caused.
MR2RP is a delay-oriented multi-rate/multi-range routing protocol for IEEE 802.11
ad hoc networks [68]. It is designed to maximize the channel utilization and minimize the
network transfer delay. The medium access control (MAC) protocol is analyzed to predict
the transfer delay of a routing path. The authors assume: (a) the packet arrival process is
a Poisson process, (b) all nodes have the same packet arrival rate, (c) each node knows
the buffer information of every other node, (d) every node knows the connectivity matrix
of the network so that the Dijkstra algorithm can be employed to find the shortest path.
SAGA is based on weaker assumptions as discussed in section 5.3.1. It will be preferable
if the delay is estimated locally without exchanging information among neighbors.
Quality-of-Service (QoS) routing protocols for ad hoc networks select routes with suf-
ficient resources to satisfy certain requirements such as delay or bandwidth [69–71]. They
work on a per-connection basis. The QoS routing requires the underlying MAC protocol to
support and guarantee resource reservation as well as provide information and constraints
about delay and bandwidth, etc. If QoS support is not available, SAGA’s delay estima-
tion methods can be extended for contention-based media access protocols to provide this
information to the upper layer protocols and applications.
86
5.7 Conclusion
Congestion control can be a problem in ad hoc networks. Compared to the traditional
solutions at the transport layer, SAGA routing protocol is implemented at the network
layer. SAGA protocol integrates the channel spatial reuse with the multi-hop nature of ad
hoc routing to reduce congestion. SAGA is a distance vector routing protocol that uses
intermediate delay (IMD) instead of hop count as the distance. The use of IMD enables
routing protocols to select routes that bypass hot spots where contention is intense. The
lazy route query operation in SAGA protocol uses a special route advertisement for route
discovery. Multiple queries can be included in one advertisement packet to accelerate the
establishment of needed routes. SAGA provides an approach to reduce the oscillation of
the value of IMD and makes the routes stable.
The use of IMD in routing decisions can enhance the performance of many routing
protocols. It is especially of benefit to networks where topology changes are much less
frequent than traffic changes. The lazy route query can be applied to other proactive
routing protocols that do not have a dedicated route discovery operation. SAGA proto-
col reduces congestion at every intermediate node. It can be used as a complementary
scheme to the end-to-end congestion control/avoidance mechanisms. The proposed delay
estimation methods can be extended for contention-based media access protocols to pro-
vide quality of service (QoS) information to upper layer protocols and applications. The
intermediate delay obtained from SAGA protocol can be used to improve the accuracy of
round-trip-time (RTT) estimation for TCP connections.
This research provides methods to estimate the delay at a node using only local in-
formation. When a node has recent traffic, statistical methods are used to evaluate the
mean of the delay. Otherwise, the underlying MAC protocol is analyzed and probability
methods are applied to compute the expectation of the delay. We analyze the packet trans-
mission procedure of the distributed coordination function in the IEEE 802.11 standard
as a case of the practical study. These methods are applicable to other contention-based
media access protocols.
87
A series of experiments have been conducted to study the performance of routing pro-
tocols under congestion. Two types of UDP traffic as well as the TCP traffic are considered
and the offered traffic load is taken as the input parameter. The maximum moving speed
of nodes and the number of connections are varied. SAGA performs better than DSDV
in all our measurements. A summary of comparison of SAGA with AODV and DSR for
throughput, overhead, and end-to-end delay is as follows.
• SAGA is able to deliver around 90% of the data packets with an offered traffic load
up to 300 kb/s. It can offer a peak throughput of 370 to 400 kb/s in all experiments.
This is 1.5 to 3.5 times as compared to the throughput of AODV and DSR.
• Overhead is measured as the ratio of the routing load to the data successfully deliv-
ered to the destination. The overhead of SAGA remains in a range of 15% to 50%.
In similar cases, the overhead of AODV and DSR varies widely and increases fast
as the offered traffic load goes high. The overhead of SAGA is as low as 10% of
that of AODV and 12% of that of DSR in high traffic load.
• For low traffic load, the average end-to-end delay of SAGA is the same as that of
AODV and DSR. When traffic reaches 500 kb/s, the delay of SAGA is 50% less
than that of AODV and 80% less than that of DSR.
Evaluating SAGA protocol in an emulation instead of simulation environment is prefer-
able for its success in real world use. In the future, we plan to use the mobile ad hoc emu-
lator MobiEmu [72] to conduct experimental studies. The impact of the accuracy of delay
estimation on the performance of SAGA protocol will be investigated. The results of the
research on the lifetime of routes in mobile ad hoc networks [73] will be adopted to im-
prove the accuracy of delay estimation. Research will be conducted to integrate SAGA’s
congestion reduction mechanism with the TCP congestion control algorithms. The idea
of randomization [64] may be adjusted for SAGA protocol to decrease routing overhead
and provide better congestion reduction.
88
6 HIERARCHICAL ARCHITECTURE FOR SUPPORTING MOVABLE BASE
STATIONS IN WIRELESS NETWORKS
6.1 Introduction
Wireless LAN is significant for people to keep connected on the move. Stationary
sites (i.e., base stations) provide high-speed network connections for mobile hosts. For
instance, IEEE 802.11a supports up to 54 Mbit/s communication capacity [74]. The fixed
infrastructure makes it easy to manage the network, enforce security policies, and extend
the system. It, however, limits the deployment of the network in environments where
wireless access to a wired backbone is either inefficient or impossible. For tactical military
networks, the fixed base stations are attractive targets, therefore highly vulnerable.
Most limitations of wireless LAN, such as inflexibility and vulnerability, can be elim-
inated by letting base stations move. Base on this idea, we propose a new type of wire-
less networks calledwireless network with movable base stations(WNMBS). WNMBS is
comprised of mobile hosts andmovable base stations(MBS). It can be rapidly deployed
without any preexisting infrastructure. Flexibility can be achieved without losing much
scalability. Supporting movable base stations in wireless networks introduces a lot of
challenging research questions. One fundamental problem that requires investigation is
how to organize MBS and effectively maintain the dynamic network topology. Because
all base stations and mobile hosts are moving, the location of a host is not determinable by
its network address. Traditional routing protocols for wireless LAN are not suitable in this
circumstance. The ad hoc routing protocols do not scale well, as indicated in [75]. They
do not take advantages of movable base stations either. Thus, the design of a new routing
protocol is mandatory.
We propose a hierarchical structure to support movable base stations in wireless net-
works and address the issues of network maintenance and routing. This architecture is
89
called hierarchical mobile wireless network (HMWN). The rest of this chapter is organized
as follows. Section 6.2 discusses the design considerations. The network architecture and
four basic operations are described in Section 6.3. Section 6.4 presents the detail of an effi-
cient membership management protocol. The segmented membership-base group routing
protocol is proposed in section 6.5. In Section 6.6, a simulation evaluation and its result
are discussed. Section 6.7 discusses related work. Section 6.8 concludes the chapter.
6.2 Design considerations
WNMBS has its unique characteristics that need to be considered in the design of the
network architecture. The following issues have been taken into account.
6.2.1 Asymmetric capacity and asymmetric responsibility
Most mobile hosts are portable computing facilities such as PDA, GPS, notebook com-
puter, etc., with portable wireless communication devices. These facilities have limited
system resources and low computing capabilities. Lightweight batteries may power these
facilities along with their communication devices. The weak power and the limited battery
life will impose restrictions on the transmission range, communication activity, and com-
putational power of the communication devices. Such mobile hosts can hardly afford the
overheads of providing network services. On the other hand, movable base stations (e.g.,
workstations mounted on vehicles) are powered by heavy-duty batteries, equipped with
high-speed communication devices. They are capable of providing reliable network ser-
vices. The design of the network architecture should fully utilize the capacity of movable
base stations and minimize computation and communication overheads for less power-
ful mobile hosts. For instance, computation-complex and resource-consuming operations,
such as routing maintenance and authentication, are done at MBS.
90
6.2.2 Coordinated movement
The random way-point mobility model [9] is commonly used to generate the move-
ment of mobile hosts in the study of ad hoc networks. According to this model, individuals
move independently. The speed and direction of the motion in the new time interval have
no relation to those of the motion in the previous time interval. In reality, the members
belonging to a group tend to coordinate their movements. The reference point group mo-
bility (RPGM) model [76] describes this kind of movement. RPGM partitions the network
into several groups. Each group has a logical center. The center’s motion defines the mo-
tion of the entire group. Each member in a group has independent random motion with
respect to the logical center in addition to the group’s motion.
6.2.3 Localized traffic
The reality of network traffic is that a small percentage of hosts in a domain are com-
municating outside of the domain at any given time. Many (if not most) hosts never
communicate outside of their domain [77]. For example, it is much more likely that com-
munication will take place between two soldiers in the same battalion, rather than between
two soldiers in two different brigades. To take advantage of this kind of traffic pattern, the
design of networks should give priority to intra-domain communications.
6.2.4 Heterogeneous wireless networks
In large scale applications, incompatible wireless networks, such as bluetooth net-
works, waveLAN networks, or satellite networks, may coexist. A desirable feature of the
network architecture is the capability of accommodating heterogeneous wireless networks
and providing simultaneous and seamless support for different MAC protocols. MBS that
are equipped with multiple wireless network interfaces are needed to forward packets be-
tween two groups that use incompatible protocols (like routers in wired networks).
91
6.3 Network architecture
Based on the considerations discussed in the previous section,hierarchical mobile
wireless network(HMWN) is designed to support WNMBS. It can be applied to ad hoc
networks as well to build a virtual hierarchy. To broaden its application, HMWN is pre-
sented in the following sections in a generic way, in which movable base stations are
treated as a special type of mobile hosts.
6.3.1 Definitions
The following is a set of definitions that will be used in the rest of the chapter.
Definition 1: A group is a set of mobile hosts. Each group has one representative
(i.e., agent). A group is denoted asgroup(M), whereM is the agent. A host can be an
agent for at most one group. The home group (HG) is where the mobile host registers
its membership. A foreign group (FG) is a group other than theHG. The current group
(CG) is the one to which the host currently attached. The corresponding group agents are
called home group agent (HGA), foreign group agent (FGA), and current group agent
(CGA), respectively. Usually, movable base stations are chosen to be agents.
For every mobile host, itsHG is assigned by the “Grouping” operation.This relation-
ship keeps unchanged during the life-time of the network. A mobile host’sCG is changed
when the “Migration” operation completes.
Definition 2: The groups in a HMWN system form a group hierarchy. The level of a
groupG, which is denoted aslv(G), represents how close it is to the root of the hierarchy.
The lower the level is, the closer the group is to the root. The level of the root group is 0.
Any mobile host can be a member of two different groups, in one of which it is the
agent, in the other one it is a non-agent member. Suppose the agent of groupG1 is a non-
agent member of groupG2, thenlv(G1) = lv(G2) + 1. If a mobile hostMH is a member
of groupG, the level ofMH is
lv(MH) =
lv(G), MH is the agent of groupG;
lv(G) + 1, otherwise.(6.1)
92
Definition 3: A groupG1 is a subgroup of groupG2 if and only if
1. the agent ofG1 is a non-agent member ofG2
2. or the agent ofG1 is a non-agent member of one ofG2’s subgroups.
G2 is called a supergroup ofG1. Operatorssub(G1, G2) and sup(G2, G1) are used to
denote thatG1 is a subgroup ofG2 andG2 is a supergroup ofG1, respectively. In HMWN,
sub andsup are partial orders.
Definition 4:A domain derived from a groupG consists of and only consists ofG and
all its subgroups, denoted asdomain(G). The group agent ofG is also the domain agent
of domain(G). Derived domains have the following property.
domain(G1) ⊆ domain(G2) ⇐⇒ sub(G1, G2) (6.2)
Definition 5:A closure domain of two groupsG1 andG2, denoted asclosure(G1, G2),
is the smallest derived domain that containsG1 andG2. Formally, closure(G1, G2) =
domain(G) if and only if
1. G1 ⊆ domain(G) andG2 ⊆ domain(G)
2. For any deriveddomain(G′), G1 ⊆ domain(G′) and G2 ⊆ domain(G′) =⇒domain(G) ⊆ domain(G′)
6.3.2 An example
Figure 6.1 is an example of the HMWN system. Every small square represents a
mobile host and the dark ones are group agents. A solid line between two mobile hosts
represents the wireless link. The dashed line circles represent groups and the solid line cir-
cles represent derived domains. The root group only contains three members{A,B,C},
whereA is the agent. There are two level 1 groups,{B,D,E} and{C,F,G}. B andC are
group agents, respectively.D, E, F , andG are agents for level 2 groups. Figure 6.2 shows
an alternate representation of the group hierarchy, where every group is represented by its
93
A
B
C
D E
FG
x
yz
t
s
domain(A)
domain(B)
domain(C)
Figure 6.1. Hierarchical mobile wireless network
agent at a lower level. In this network, thedomain(A) contains 7 groups and all hosts in
the system. Thedomain(B) consists of 3 groups and mobile hosts{B,D,E, s, t, x, y, z}.
In HMWN, mobile hosts that belong to the same group use a multi-hop ad hoc routing
protocol to communicate. Communication with a host outside the group is accomplished
by the segmented membership-based group routing protocol presented in Section 6.5.
6.3.3 Basic operations
The following four basic operations are defined for setting up and maintaining a HMWN
system.
1) Groupingis the operation used to set up the static membership in a HMWN system.
It is only performed at the bootstrapping phase. “Grouping” is accomplished in two steps.
94
level 0 group
level 2 groups
level 1 groups
B
B
C
C
A
E
E
D
D
F
F
G
G
Figure 6.2. Hierarchy of groups
95
The first is to organize mobile hosts into groups (i.e., assignHG for each mobile host).
The second is to determine group agents (HGA). The criteria for “Grouping” include
• Mobility: If a set of mobile hosts are going to coordinate their movements, they may
form a group.
• Organization: If all mobile hosts belong to a organization that has a well established
hierarchy, the hosts can be grouped based on this hierarchy.
• Wireless MAC protocol: If multiple wireless MAC protocols are used in the net-
work, the mobile hosts that support compatible protocols may be grouped together.
• Capacity: Capacity is used to determine group agents. The higher the capacity is, the
greater the chance is that the mobile host will be chosen as an agent. Several factors
are taken into consideration when the capacity of a mobile host is evaluated, e.g., the
computation capability, system resource, power level, communication bandwidth
and range, the number of wireless network interfaces.
This operation can be done in a distributed or centralized way.
• Mobile hosts may autonomously organize themselves into groups, then supergroups.
In the autonomous procedure, each agent will exchange the organization, the MAC
protocol, and capacity information with its neighbors to determine the static mem-
bership relationship. This process is accomplished in a distributed way. It is hard to
obtain the optimal result.
• A trusted authority may take charge of the operation. Every mobile host reports
its information to the authority. The authority employs some global optimization
algorithm to establish the hierarchy and distributes the result to all participated hosts.
The first scheme is also suitable for self-organizing ad hoc networks, in which mobile
hosts have no prior knowledge about the network. In practice, a mobile host is usually
assigned a home agent before joining the network, or knows some information that is
96
helpful for grouping. Automatically grouping in a distributed fashion itself is a non-trivial
problem. We do not address it in this dissertation.
2) Registrationis the operation that a mobile host must complete before it can connect
to the network. “Grouping” only determines the static membership. “Registration”, along
with “Leaving” and “Migration”, maintains the dynamic topology of the network (e.g.,
CG for a mobile host). Registration takes place between a mobile hostMH and itsHGA.
One-hop registration is recommended to reduce the possibility of denial-of-service and
man-in-the-middle attacks.
This operation begins withMH broadcasting the“Registration” request. If theHGA
is within the neighborhood, the operation continues with an identity verification process.
Upon successfully registered,MH will obtain the group information such as group ID,
group shared secrets, etc. from theHGA, and set theHGA to be itsCGA. In case that
MH itself is an agent of another group, all hosts in the deriveddomain(MH) implicitly
become members of the network.MH keeps moving and sending out the request periodi-
cally if it cannot reach theHGA directly. Other hosts may provide aid to locate theHGA
so thatMH can adjust its movement.
If connectivity rather than security is preferred, remote registration (i.e.,MH registers
itself to theHGA via intermediate hosts) will be allowed.
3) Leavingoperation is completed by group agents. It may be triggered by two events.
• When a mobile hostMH decides to leave the network (along with all hosts in the
deriveddomain(MH)), it sends a ”leave group” message to itsCGA.
• When the agent finds out that the route to a mobile hostMH is broken, it starts a
Leaving Timer. If a route toMH cannot be reestablished or a “Migration”message
has not received within theLeaving Interval Timeas described in equation 6.3, the
agent starts the “Leaving” operation.
Leaving Interval Time
= Robustness × Ad Interval × (Max Hop + 1) (6.3)
97
A
B
C
D E
FG
x
y z t
s
domain(A)
domain(B)
domain(C)
Figure 6.3. After migration
TheAd Interval is the time interval between the route advertisements sent out by a
host. TheMax Hop is the hop number of the longest route in the agent’s routing
table.Leaving Interval Timeis the maximum time it will take to getMH ’s routing
information if MH is still a member of the group. TheRobustnessallows tuning
for the expected packet loss on wireless links. The “Leaving” operation is able to
tolerate (Robustness − 1) failures. ThusRobustnessmust be greater than 1. If the
system is expected to be lossy, theRobustnessmay have a larger value.
After theCGA of MH updates the membership information, it will forward the ”leave
group” message to its ownCGA.
4) Migration operation is initiated by a mobile host that decides to leave its current
group and join a foreign group. Usually, when a hostMH realizes that theCGA is no
longer reachable, it starts this operation by sending out a “Migration” request. Foreign
98
agents that are in the neighborhood reply this request based on the MAC protocol compat-
ibility and capacity, and the security policy that determines whether or not to provide the
migration support.MH chooses theFGA whose reply comes first, sets it to be theCGA,
and invokes the hand-off procedure. Every agent that replies the request will start a timer.
When the timer expires, the agent will cancel the operation.
Figure 6.3 illustrates the topology of the example HMWN system shown in Figure 6.1
after mobile hostz migrated fromgroup(D) to group(E).
6.4 Membership management
Maintaining the network topology in an efficient way is significant in a HMWN sys-
tem. Essentially, it is a membership management problem because the mobile hosts are
organized as hierarchical groups. The following subsections present the membership man-
agement protocol.
6.4.1 Data structure
The membership information is mainly used for two purposes. The first is to verify the
identity of a host (i.e., the static membership). The second is to help routing protocols to
choose the proper route to forward packets (i.e., the dynamic membership). Each agentG
maintains two separate tables.
StaticMemberTable contains the identification information of mobile hosts whose
HGA is G. This table is mainly used by security protocols such as authentication and
identity verification. The table has an entry for every potential member, which is a 3-
tuple {ID, sharedsecret, public key}. Initially, an entry only contains theID and the
sharedsecret. After registration, the public key of the member will be recorded in the
entry.
Current MemberTablecontains the information of all the mobile hosts that currently
belong to the domaindomain(G). The entry of the table is a 3-tuple{ID, intermedi-
ate host, homeagent}. The intermediatehost is the non-agent member in this group
99
whoseCurrent MemberTablealso contains the mobile host (i.e., the mobile host is in the
domain derived from theintermediatehost). Thehomeagentis theHGA of the mobile
host. This table is used by the routing protocol to locate mobile hosts.
Depend on the size of the tables and the available memory, these two tables can be
maintained using a hash table, a ordered list, or a trie to accelerate the searching process.
“Registration”,“Leaving”, and “Migration” will operate on these two tables.
6.4.2 Registration
Upon successful registration, a host will get the group information from the agent.
The host sets the agent to be itsCGA. In case that security protocols are deployed, a
mutual challenge-and-response process will be initiated to verify the identity of the host
and the agent. If verification succeeds, the agent will record the host’s public key in the
corresponding entry ofStaticMemberTable, the host will get the group key, the agent’s
public key, and other information required by the security protocols such as a certificate.
The host will send a list of all members in itsCurrent MemberTable to the agent
so that all members in its derived domain will be implicitly registered. This list will be
forwarded via the path from the agent to the root of the hierarchy. Every agent on the path
will add the members to its ownCurrent MemberTable.
6.4.3 Leaving
When a host leaves a group, all members in its derived domain also leave the group
implicitly. The host sends a list of all members in itsCurrent MemberTableto the agent.
This list will be forwarded via the path from the agent to the root of the hierarchy. Every
agent on the path will remove the members from its ownCurrent MemberTable.
100
6.4.4 Migration
When a mobile hostMH is leaving the current groupG1 and joining another group
G2, both theCGA and theFGA will update theirCurrent MemberTable. If MH is an
agent, all mobile hosts indomain(MH) also implicitly leavedomain(G1) and join the
domain(G2). After joining the foreign group,MH will send messages to theFGA and
theCGA to help them update the membership.
At theFGA side,MH sends the following message to the foreign agent.
[ADD, ID, previous agent,member list]
whereID is the identification ofMH, previousagentis MH ’s CGA before joining the
group,memberlist is MH ’s Current MemberTable.
For each host in thememberlist, theFGA adds it to theCurrent MemberTableif it
does not exist already, and sets theintermediatehostto theMH that sent the message. If
previousagentis not a member in theCurrent MemberTable, theFGA sends the same
message to its ownCGA. Every agent that receives the message will update the member-
ship as well.
At theCGA side,MH sends the following message to the current agent.
[REMOV E, ID, foreign agent,member list]
whereID is the identification ofMH, foreign agent is the agent of the foreign group,
memberlist is MH ’s Current MemberTable.
If the foreign agentis also a member in theCurrent MemberTable, which means the
MH moves from one sub-group to another, then theCGA does nothing. Otherwise,
it removes every host in thememberlist from theCurrent MemberTableand forwards
the message to its ownCGA. Every agent that receives the message will update the
membership as well.
Figure 6.4 shows the difference between “Registration”, “Leaving”, and “Migration”
operations with respect to the modification ofCurrent MemberTable. The small circles
represent the mobile host. For “Registration” and “Leaving”, the effect will be prop-
agated to the root of the hierarchy. Thuslv(A) + 1 unicast are required, whereA is
101
+
+
+ -
-
-
-
Registration Leaving Changing Membership
+
Remove members from Current_Member_Table
Add members to Current_Member_Table
-
+
Figure 6.4. Membership modification
the agent. For “Migration”, the effect is only propagated to the agent of the domain
closure(previous agent, foreign agent). The number of required unicast is
lv(previous agent) + lv(foreign agent)
−2 ∗ lv(closure(previous agent, foreign agent)) (6.4)
6.5 Segmented membership-based group routing
Segmented membership-based group routing (SMGR) protocol is proposed for the
HMWN system to take advantage of the hierarchical group structure and available mem-
bership information.
6.5.1 Data structure
SMGR protocol requires two tables. One is the routing table, in which each entry is
a 4-tuple<destination, nexthop, distance, sequencenumber>. The sequencenumber
represents the freshness of the route. Each host maintains asequencenumberfor itself.
102
D D 1 30
E E 1 23
A A 1 22
C C 1 20
Routing Table of B
s E
t E
x D
y D
z E
D D
E E
G C
Membership Table of B
Figure 6.5. Membership table and routing table
This number is monotonically increasing. Only routes to the group-mates are maintained
in the routing table. These routes are updated using DSDV [1] protocol.
The other is the membership table, in which every entry is a 3-tuple<final destination,
intermediatehost, routing entry>. routing entryis a pointer to the entry in the routing ta-
ble that specifies the route to theintermediatehost. Every entry inCurrent MemberTable
has a corresponding entry in this table.
Take host B in Figure 6.3 as an example, Figure 6.5 shows the routing table, the mem-
bership table, and the pointers maintained by B.
The size of the routing table is bounded by the size of the group, which is nearly a
constant.
SMGR protocol will add a header, which is a 4-tuple<source, final destination, inter-
mediatehost, nexthop>, to each packet. The header is used to route the packet.
6.5.2 Routing
When a host receives a data packet, either from another host or from a application
running on itself, it takes different actions to forward the packet, based on whether it is the
intermediatehostor not. Here we assume that the routing table is up-to-date.
103
If the host is not theintermediatehost, it simply forwards the packet based on the
available routing information. Otherwise, it is responsible for locating the nextintermedi-
ate host(or the final destination) from its membership table. The packet is forwarded to
the nextintermediatehostif it is located, otherwise, the packet is forwarded to theCGA.
Since the root group agent can locate any mobile host, the packet will eventually reach the
destination. In the routing process, “membership expires” or “redirect” message may be
sent out to update the membership information.
A host will removes the corresponding entry from the membership table when it re-
ceives a “membership expires” message. When a host receives a “redirect” message, it
adds an entry in the membership table, setintermediatehostto be the redirected host.
Figure 6.6 shows the pseudo code of the SMGR routing algorithm.
6.6 Evaluation
A simplified version of SMGR has been implemented in the network simulator ns2 [18].
In this version, the membership modification is completed through broadcast instead of
unicast. It is predictable that more protocol overhead will be introduced by the simplifi-
cation. We have also implemented the computation delay component to simulate different
computation capacities. The purpose of the experiment is to evaluate the scalability of
HMWN in terms of protocol overhead. Because there is no other routing protocol de-
signed for WNMBS, we apply HMWN to ad hoc networks for comparison purpose. Since
SMGR utilizes distance vector, we compare it with two distance vector based ad hoc rout-
ing protocols, DSDV and AODV.
In this experimental study, we take the protocol overhead (protocol load divided by
throughput) [8] as the metric to evaluate the scalability of routing protocols. The exper-
iments simulate a 1000m x 1000m area. Random way-point mobility model is used to
generate movement for mobile hosts, the maximum speed is 5m/s, the pause time is 3
seconds. The number of end-to-end connections is equal to the number of hosts. The
source-destination (S-D) pair of each connection is randomly chosen. Constant bit rate
104
if it is the final_destinationsend the packet to the corresponding application;
else if it is the next_hopfind out the route to the intermediate_host;change next_hop and send out the packet;
else if it is the intermediate_hostsearch the Current_Member_Table;if an entry e exists for the final_destination
set the intermediate_host to e.intermediate_host;get the routing table entry re;set the next_hop to re.next_hop;send out the packet;if the packet comes from a host which is in thesame group of e.intermediate_host
send a "redirect" message to the host;else
if the packet comes from a host of which it isthe agent
set the intermediate_host to CGA;send the packet to CGA;
elsesend out a "membership expires" message tothe source;
else if it is the sourcesearch the Current_Member_Table;if an entry e exists for the final_destination
set the intermediate_host to e.intermediate_host;get the routing table entry re;set the next_hop to re.next_hop;send out the packet;
else if it is not the root of the hierarchyset the intermediate_host to CGA;send the packet to CGA;
elsedrop the packet and notify the application;
elsedrop the packet silently;
Figure 6.6. SMGR routing algorithm
(CBR) traffic is generated for all connections. The number of hosts ranges over{20, 30,
40, 50, 60}. For each value, five scenarios are created. Individual simulation runs 1000
seconds. The protocol overhead is computed from the traffic trace file.
The result of the experiment is shown in Figure 6.7. The curves present the mean
value of the protocol overhead for each protocol. When the number of hosts is less than
40, three protocols have similar performance, with AODV being outperformed a little bit.
When the number of hosts reaches 60, the overhead of DSDV is about 50% higher than
105
20 25 30 35 40 45 50 55 600
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Nor
mal
ized
Pro
toco
l Ove
rhea
d
Number of Mobile Host
simple SMGRDSDVAODV
Figure 6.7. Protocol overhead versus number of mobile hosts
that of the simple SMGR, while the overhead of AODV is about 38% higher. The result
shows that the simple SMGR is more scalable in terms of protocol overhead.
Considering the random way-point mobility model and the random traffic pattern that
are used for the experiments favor ad hoc networks, and the simple SMGR introduces extra
protocol overhead because of unnecessary broadcast, we may expect a HMWN system
supported by SMGR protocol to be more scalable with the presence of movable base
stations.
6.7 Related work
Examples of integrated heterogeneous wireless networks include the integrated ad
hoc and cellular networks. X. Wu et al. proposed mobile-assisted connection-admission
(MACA) channel allocation scheme to achieve load balancing in a cellular network [6].
In MACA, some special channels are used to connect mobile units from different cells.
106
When a mobile unit cannot connect to its own base station due to heavy load, it may be
able to get connected to its neighboring cell’s base station through a two-hop link.
A similar approach, integrated cellular and ad hoc relaying systems (iCAR), is pro-
posed by H. Wu et al. in [5]. It addresses the congestion problem due to unbalanced traffic
in a cellular system and provides interoperability for heterogeneous networks. The basic
idea is to place a number of ad hoc repaying stations at strategic locations, which can be
used to relay signals between mobile hosts and base stations.
Multihop cellular networks (MCN) is presented by Y.-D. Lin and Y.-C. Hsu in [78].
MCN allows wireless transmission to go through mobile stations in multiple hops in
the cellular networks. It reduces the number of required base stations and improves the
throughput, while limiting path vulnerability encountered in ad hoc networks.
H. Luo et al. proposed the unified cellular and ad-hoc network (UCAN) to enhance
cell throughput and maintain fairness in the third generation (3G) data networks [79]. The
scheduling algorithm for the 3G base station is refined so that the throughput gains of
active clients are distributed proportional to their average channel rate. A secure crediting
mechanism is developed to motivate users to participate in relaying packets for others.
In [80], S. Nesargi and R. Prakash present a distributed, dynamic channel allocation
(DCA) algorithm for virtual cellular networks where the fixed base stations are replaced by
mobile base stations. Principles of mutual exclusion pertaining to distributed computing
systems are employed in the development of the algorithm. This work to some extent
provides the physical layer support to our research in WNMBS.
R. Ramanathan and M. Steenstrup proposed a MMWN system, an acronym for multi-
media support for mobile wireless networks [81]. A MMWN system consists of switches
and endpoints. While both can be sources of or destinations for packets, only switches
can route packets. The switches and endpoints are organized using hierarchical clustering
to provide support for quality of service. Routing information is distributed in the form of
link states, which contains connectivity and service information pertaining to clusters at all
levels within the hierarchical control structure. The quality-of- service routing is realized
by establishing and maintaining a virtual-circuit between the source and destination.
107
S. Banerjee and S. Khuller present the design and implementation of a clustering
scheme for hierarchical control in multi-hop wireless networks in [82]. The clustering
problem is defined in a graph theoretic framework. The properties of the underlying com-
munication graphs of wireless network are exploited to achieve desired solutions, which
satisfy the requirements such as cluster connectivity, upper and lower bounds on cluster
size, low overlap between two clusters, etc.
Many research efforts [83–85] introduce structures on ad hoc networks to provide scal-
able solutions for routing, location management, and resource allocation. Most schemes
assume that ad hoc networks are self-organized to discover and maintain the structure. It
requires extra message exchanges that may consume a large portion of the limited band-
width.
6.8 Conclusion
We present a hierarchical structure to support movable base stations in wireless net-
works. In a HMWN system, mobile hosts form hierarchical groups. Group agents (usually
movable base stations) take major responsibilities for managing membership and routing
packets. HMWN integrates the routing protocol with membership management to re-
duce overhead. It is capable of accommodating incompatible wireless MAC protocols and
managing heterogenous wireless networks in a unified way. Four basic operations that are
used to set up and maintain the hierarchy have been discussed. The detail of an efficient
membership management protocol is presented. The segmented membership-base group
routing protocol for HMWN is proposed. An experimental study is carried out to com-
pare the scalability of SMGR with AODV and DSDV ad hoc routing protocols in terms of
protocol overhead. The SMGR outperforms these two protocols for about 50% when the
number of hosts reaches 60.
This work is only the first step in the research on wireless networks with movable
base stations. We are developing multiple MAC protocols and supporting modules in
ns2 to carry out experimental studies on HMWN and SMGR protocol with respect to
108
other performance metrics. Automatically grouping in a distributed way and introducing
security mechanisms are the next steps. We hope this work will help to build a foundation
for the research of flexible, scalable, and secure wireless networks.
109
7 SECURING WIRELESS NETWORKS WITH MOVABLE BASE STATIONS
7.1 Introduction
7.1.1 Wireless network with movable base stations
Wireless communication technology is significant in networking infrastructure. Mo-
bile ad hoc networks and wireless LAN are two typical packet-switching wireless net-
works1.
A mobile ad hoc network consists of mobile hosts that communicate with each other
over multi-hop wireless links in a collaborative way [86]. There is no fixed infrastructure
or stationary base station to coordinate communications. These characteristics provide
users with maximum flexibility, at the cost of limitations on scalability. The scalability
problem is analytically studied in [75]. The result shows that even the most scalable
routing protocol introduces a total overhead ofO(N1.5), whereN is the number of hosts.
The experimental study also shows that the increase of the number of hosts is the dominant
cause for performance degradation [56].
In a wireless LAN, stationary sites (i.e., base stations) provide high-speed network
connections for mobile hosts. The fixed infrastructure makes it easy to manage the net-
work, enforce security policies, and extend the system. It, however, limits the deployment
of the network in environments where wireless access to a wired backbone is either inef-
ficient or impossible. For tactical military networks, the fixed base stations are attractive
targets, therefore, highly vulnerable.
Most limitations of wireless LAN, such as inflexibility and vulnerability, can be elim-
inated by letting base stations move. We deviate from the conventional wireless networks
1Sensor network is a new class of wireless networks that has become an attractive research area. A sensornetwork is essentially an ad hoc network that consists of a large number of tiny disposable and low-powerdevices. These devices are immobile, or have low mobility as compared with hosts in mobile ad hoc net-works.
110
and propose wireless network with movable base stations (WNMBS). WNMBS is com-
prised of mobile hosts and movable base stations. The movable base stations typically are
mounted on vehicles such as tanks and trucks and form a mobile backbone. They have
more resources than mobile hosts in terms of memory, computation capability, transmis-
sion power, energy supply, etc. Neighboring base stations use wireless links to commu-
nicate. Because all base stations and mobile hosts are moving, the location of a node is
not determinable by its network address. The traditional network architecture and routing
protocols for wireless LAN are not suitable in this circumstance. We develop hierarchical
mobile wireless network (HMWN) to support WNMBS. The details of HMWN, includ-
ing the network maintenance mechanism, the routing protocol, and control overhead, are
presented in the previous chapter.
7.1.2 Security issues in WNMBS
Achieving security in a wireless network is challenging because of:
• The use of wireless channels that are susceptible to link attacks [87];
• Roaming in a hostile environment with relatively poor physical protection that makes
a mobile host vulnerable;
• Dynamic network topology and memberships.
Security mechanisms have been proposed for protecting a single wireless link, such as
secure protocols for wireless LAN [88, 89]. The use of cryptography to secure ad hoc
routing protocols has been investigated in [90–93]. A scalable security solution for mobile
ad hoc networks is proposed in [94]. The idea of threshold secret sharing and secret share
updates is used to tolerate intrusions. Ariadne is an on-demand ad hoc routing protocol
that provides security against one compromised node and arbitrary active attackers [95].
Ariadne relies only on symmetric cryptography, thus it does not require a trusted hardware
or powerful processors. These research efforts require mobile hosts to be able to identify
111
each other based on some priori knowledge. The following mechanisms are usually used
for identification. They have deficiencies when being applied to wireless networks.
• All hosts share a secret key so that everyone can prove its membership by showing
the knowledge of this secret key. This scheme is relatively insecure. If one host is
compromised, the whole system is compromised.
• Every host knows the public keys of all other hosts so that it can identify a host by
using public-key cryptography. This scheme is not scalable. It requires all hosts to
be known before the network is set up. If a host wants to change its public/private
key pair, it has to inform all others in the system.
• There exists a centralized trusted entity, such as a key distribution center (KDC) or a
trusted third party (TTP), which knows the public key of every host. Two hosts can
use some authentication protocol, such as Yahalom, DASS, Woo-Lam, etc. [96], to
authenticate each other. In this scheme, the centralized entity is the bottleneck of
a system that will decrease the effectiveness of security solutions. It is prone to
denial-of-service attacks and may become the single point of failure.
In a WNMBS, the mobile backbone (i.e., base stations) is typically maintained by
system administrators (e.g., service providers) and provides network services to mobile
users. The base stations, with appropriate security enhancements, naturally form a dis-
tributed trusted entity that is capable of balancing service load and tolerating site failures.
To utilize movable base stations as a distributed trusted entity, research questions, such
as how to organize base stations, how to distribute keys, and how to authenticate mobile
hosts, need investigation.
We present mechanisms integrated with HMWN to secure WNMBS. The protection
of network infrastructure, authentication and key distribution, and secure roaming support
are addressed. The rest of the chapter is organized as follows. Section 7.2 discusses the
security objective and assumptions. Secure packet forwarding mechanism that protects the
network infrastructure is proposed in section 7.3. Section 7.4 presents the authentication
protocol. Section 7.5 discusses the secure roaming support. The computation overhead of
112
the security mechanisms is numerically investigated in Section 7.6. Section 7.7 concludes
the chapter.
7.2 Security objective and assumptions
We focus on protecting the network infrastructure against both passive and active at-
tacks, such as insertion, modification or replay of control messages, and traffic analysis.
End-to-end data communications are protected from unauthorized access by higher layer
protocols. As long as the network infrastructure is available and secure, the two ends of
a communication can always set up a symmetric secret key by using some key-exchange
algorithm such as Diffie-Hellman or COMSET [96]. The data packets can be encrypted
by using the secret key to ensure confidentiality and integrity.
The objective is achieved by deploying secure packet forwarding and authentication
protocols that are presented in the following sections. These security mechanisms are
based upon the following assumptions:
• The wireless communication is robust with respect to attacks against the physical
layer. These layers are well protected by lower-layer mechanisms, such as anti-
jamming techniques [97,98].
• The underlying cryptography primitives, such as digital signature and encryption,
are practically secure (i.e., they are unbreakable with current computation power).
• All base stations know each other’s public key (For instance, if each group has 50
members, a 5000-node networks requires about 100 base stations to maintains about
150 public keys, instead of 5000 nodes, most of which are resource-poor mobile
hosts, to maintain 5000 public keys.).
7.3 Protection of network infrastructure
Unlike a wired network where the infrastructure is protected by physically securing the
cables, the infrastructure of a wireless network is protected by ensuring that every mobile
113
host has correct knowledge about the current network topology and the memberships. A
mobile host obtains this knowledge by securely exchanging control information, such as
neighbors, routes, etc., with other trustworthy hosts. An adversary should not be able
to eavesdrop, insert, or modify the information. It is guaranteed by using unforgeable
encryptions.
In addition to routing and control messages, packet headers need to be encrypted.
Although encryption hides the content of a message, the packet header that contains the
source, the destination, and the next hop will expose the relationships among the involved
hosts. This is a reason why eavesdropping technology such as Carnivore is useful even
in the presence of unbreakable communication [99]. Preferred targets can be identified in
this way and attacks can be concentrated on the nerve centers. Encrypting packet headers
will effectively obfuscate relationships among hosts.
We assume that each mobile host in a HMWN system has a public/private key pair
and group members know the public key of the group agent. Each group agent maintains
a potential member list (defined by the “Grouping” operation), which contains the public
keys of mobile hosts that might be a member of that group.
The secure packet forwarding algorithm is designed for the protection of the network
infrastructure. To use a symmetric cipher, each group has a group-shared secret key. This
key is maintained and distributed by the group agent. It is renewed periodically, when a
mobile host joins or leaves the group, or at the time a compromised host is discovered.
When a mobile host X registers to a group, it authenticates itself with the group agent
and gets the group shared key K by invoking the protocol presented in Section 7.4. X uses
K to communicate with other group members confidentially. A group agent may know
two groups’ shared keys.
The pseudo-code in Figure 7.1 shows how X handles (sends, receives, and forwards)
packets after joining the group. This algorithm integrates with the routing protocol to
realize secure packet forwarding.
Encrypting and checking headers when sending, receiving, or forwarding packets
serve the following purposes.
114
Part I: sending a packet P:
1. X uses K to encrypt the header2. if P is a routing or control packet3. it uses K to encrypt the body of P4. X transmits encrypted packet P
Part II: receiving a packet P:
1. X decrypts and checks the header2. if X itself is the destination and P is a control packet3. it decrypts the body4. else5. X makes any necessary modifications to the header6. if X is a group agent AND P is sent from one group to another7. it encrypts the header with the destination group’s key K’8. if P is a routing or control packet9. it decrypts the body with K and re-encrypts it with K’10. else11. X encrypts the header with K12. X forwards P to the next hop
Figure 7.1. Secure packet forwarding algorithm
1. The correctly encrypted header testifies that a packet is sent by a member of the
group. Adversaries cannot produce such a header because they do not know the
secret key. It prevents the network from being flooded with false control and data
packets generated by malicious hosts.
2. The encrypted header ensures that routing and location information, which is valu-
able to attackers, will not be disclosed. For example, if an adversary captures a
packet and knows the next hop is host X, he can tell that X is within the radio range
of the sender and initiates attacks against X.
115
7.4 Authentication and key exchange
The capability of a mobile host to authenticate itself and obtain the group-shared key
is the basis of secure packet forwarding. In this section, we discuss the authentication and
key exchange protocol.
7.4.1 Notations and protocol
We introduce the following notations.
• X, Y: mobile hosts
• G: group agent
• gid: group ID
• R: request. It could be a request for joining a group or a request for secure roaming
support.
• T: time stamp
• K: shared secret key
• KX : public key of host X
• M: message
• EX(M): encrypting message M with host X’s public key so that only X can read M
• SX(M): signing message M with X’s private key so that every host that knows X’s
public key can verify that M is signed by X
• VX(M): verifying message M with X’s public key
• EK(M): encrypting message M with secret key K
• DK(M): decrypting message M with secret key K
116
1. X→G: <gid, X, R, SX(gid, X, R)>2. G: VX(gid, X, R)3. G→X: <gid, G, X, R, EX(gid, G, X, R, K, SG(gid, G, X, R, K))>4. X: VG(gid, G, X, R, K)5. X→G: <X, G, EK(X, G, R)>
Figure 7.2. Authentication and key exchange protocol
The protocol shown in Figure 7.2 illustrates the process invoked by the “Registration”
operation when host X joins a group whose ID is “gid”. This protocol does not use a time
stamp to guarantee the freshness of the request because a mobile host only registers once
in the network. The agent can tell if the request is new by examining the membership
information it maintains.
7.4.2 Correctness
The correctness of the protocol can be proven by adopting the logic of authentica-
tion [100]. The following terms are used.
• X believesP: host X thinks that a statement P is true.
• X seesP: host X receives a statement P.
• X controlsP: host X is trusted in the matter of the statement P.
• fresh(P): P is a fresh statement.
The following three deductions are used in the proof.
1. X sees SY (P) and X believes fresh(P)⇒ X believesY believes P.
2. X believesY believes Pand X believesY controls P⇒ X believes P.
3. X sees EK(P) and X believesX and Y share Kand X believesY controls P⇒ Y
believes K.
117
Theorem: The authentication and key exchange protocol shown in Figure 7.2 authenticates
X and G and establishes a shared key between X and G.
Proof: The following believes are held before the protocol starts.
1. G believesX controls R(because G knows that X initiates the request)
2. X believesG controls K(because X knows that G generates the shared key).
3. Both X and G believe fresh(R) (because X is yet a member of the group)
After step 2:
• G believesX believes R
• G believes R
After step 4:
• X believes fresh(K)
• X believesG believes K
• X believes K
After step 5:
• G believesX believes K
At the end of the protocol:
• X believes K
• X believesG believes K
• G believes K
• G believesX believes K
A detailed derivation of the proof is presented in [101].
118
7.4.3 Security discussion
A security protocol should be robust against malicious attacks. The authentication and
key exchange protocol is immunized to the “man-in-the-middle” attack. An adversary
can not modify the request or response because of the use of asymmetric cryptography.
The “replay” attack will not work either since this protocol is invoked only once for each
mobile host. Both X and G are capable of telling whether the request is brand new with
respect to X.
The most severe threat to this protocol is that an attacker could use it to initiate denial-
of-service (DoS) attacks against group agents. Because the mobile host does not know
the shared key and can not encrypt the packet header at this time, an attacker can discover
the identity of a group agent and locate its position by eavesdropping these requests and
analyzing the packet headers. This threat may be avoided by encrypting the packet header
of the request with the agent’s public key and the packet header of the response with the
mobile host’s public key. An attacker could not distinguish the authentication protocol
packets with other control or data packets. Furthermore, the movement of a group agent
makes it complicated for an attacker to launch continuous DoS attacks.
7.5 Secure roaming support
A mobile network allows mobile hosts to roam within the network. In wired environ-
ments, Mobile IP is the most widely used protocol to support roaming. Mobile IP is not an
ideal solution for HMWN, because (a) it establishes a “tunnel” between the home agent
and foreign agent, which consumes wireless bandwidth; (b) it does not support “group
roaming” (i.e., a whole group moves from one place to another). The essence of roam-
ing support is relocating a mobile host. SMGR protocol naturally supports roaming as it
dynamically locates the destination when forwarding a packet.
In case secure packet forwarding is required by the foreign group, the mobile host
must authenticate itself to the foreign group agent and obtain the shared key before it can
communicate with other hosts in the foreign group. This process is call secure roaming.
119
Mobile host:
1. if homeless2. broadcasts a “join a group temporarily” request3. if a response from a FGA is received4. invokes the authentication process with that agent5. if authenticated6. changes the group ID and the shared key along with the CG and CGA
Group agent:
1. if a “join temporarily” request is received2. if the security policy allows hosting3. sends a response to the mobile host4. invokes the authentication process5. if authentication succeeds6. issues a new shared key7. distributes the new key to the current group members8. sends the group information (gid, key) to the mobile host
Figure 7.3. Secure roaming support algorithm
1. X→FGA: <X, FGA, HGA, R, T, SX(X, FGA, HGA, R, T)>2. FGA→HGA: <X, FGA, HGA, R, T, SX(X, FGA, HGA, R, T)>3. HGA→FGA: <SHGA(X, KX , R, T), SHGA(FGA, KFGA, R, T)>4. FGA→X: <SHGA(FGA, KFGA, R, T), EX(FGA, X, R, T, K, SFGA(FGA, X,
R, T, K))>5. X→FGA: <X, FGA, T, EK(X, FGA, T)>
Figure 7.4. Mutual authentication protocol
7.5.1 Secure roaming support algorithm
The pseudo-code in Figure 7.3 shows the sketch of the secure roaming support al-
gorithm. This algorithm is a part of the “Migration” operation. Its purpose is to verify
120
the identity of the mobile host and distribute the shared key safely. Other issues related
to “Migration” are discussed in the previous chapter, including when to initiates the op-
eration, how to choose a foreign group to join, how to update membership, and how to
maintain routing table.
7.5.2 Mutual authentication between a mobile host and a FGA
Mutual authentication is required by secure roaming support algorithm to protect the
foreign group as well as the mobile host. Figure 7.4 shows the mutual authentication
protocol. Only messages exchanges are presented. The verifications at X, HGA, and FGA
are omitted without losing the essence of the protocol. Through this protocol, X and FGA
can get each other’s public key, which is signed by the HGA. FGA can verify that the
request is initiated by X. The fourth step ensures that only X can get K. X must verify that
K is generated by FGA using FGA’s public key. Because roaming support may be required
by the same mobile host multiple times, a time stamp is associated with each request to
prove its freshness. The use of time stamp may avoid the “replay” attack. It requires a
loose synchronization among all mobile hosts.
The correctness of the mutual authentication protocol can be proven using the logic of
authentication similarly to the proof presented in the previous section.
7.5.3 Fault-tolerant authentication
In a WNMBS, group agents are also moving. When the mutual authentication protocol
is taking place, the HGA of X may be temporarily or permanently unavailable because of
movement or failure. In this case, X’s request for the temporary membership in the foreign
group will be denied. Mobile hosts will be detached from the system if their HGAs are no
longer available. To make HMWN networks survivable from such kind of unavailability,
a fault-tolerant authentication scheme is proposed in [102].
121
In a HMWN system. A group agent itself may be a member of another group and has
its own HGA, unless it’s the root of the hierarchy. We define mobile host X’s Intention
Agent (IA) as follows:
Mobile host Y is X’s IA if and only if Y is the HGA of X’s HGA or Y is the HGA of one
of X’s IAs.
For example, in Figure 6.1, agents A and B are IAs of mobile host x. In the proposed
fault-tolerant scheme, not only its HGA, but also all its IAs know the public key of a
mobile host. A mobile host also knows all its IAs’ public keys. Each IA has a priority
based on several factors [103]. When the mutual authentication protocol fails due to the
unavailability of the HGA, the mobile host will choose the IA with the highest priority
and retry the authentication process until it is authenticated or no IA is available. With
this improvement, a mobile host at leveln can toleraten agent failures.
7.6 Computation overhead
The majority of computation overhead introduced by the security mechanisms comes
from two sources: the secure packet forwarding and the secure roaming support. We nu-
merically investigate the overhead by conducting a series of experiments and simulations.
The test-bed is a PC running Linux kernel 2.4.2. It has an Intel Celeron 700MHz CPU,
128M memory, and a 10G hard disk. Currently, even a low-end notebook computer has
better configuration than the test-bed machine in terms of computation power.
The cryptography implementations used in the experimental study are provided by the
GNU Crypto package. The testing programs are written in Java and compiled using JDK
1.3.1.
7.6.1 Overhead of secure packet forwarding
For any host that forwards a packet, it will decrypt and encrypt the packet header once
using some symmetric cryptographic algorithm. We denoteB as the bandwidth available
122
Table 7.1Encryption/decryption speed of block ciphers
Encryption Decryption CPUCipher
Speed (KB/s) Speed (KB/s) Usage
DES 4035 4061 3%
Triple-DES 1338 1323 9.8%
Twofish 1284 1277 10%
Rijndael 8185 8134 1.6%
to the mobile host,Lp as the average packet length, andLh as the length of the packet
header. We letSe be the encryption speed andSd be the decryption speed. The equation
Lh
Lp
B/Se +Lh
Lp
B/Sd (7.1)
estimates the maximum computation time required to encrypt and decrypt the data going
through the host in one second.
We take the IEEE 802.11b standard as an example, which supports up to 11Mbps
wireless bandwidth (i.e.,B=11Mbps). Suppose only the IP header is encrypted (i.e.,Lh
= 20 bytes). Based on the study of IP packet length distribution [104], we letLp = 420
bytes, the mean of IP packet length obtained from more than 200 million packets. The
computation time can be derived as follows based on equation 7.1.
Lh
Lp
B/Se +Lh
Lp
B/Sd
=20
420× 11Mbps/Se +
20
420× 11Mbps/Sd
≈ 0.0655MBps/Se + 0.0655MBps/Sd (7.2)
123
Four block ciphers are studied. They are DES (Data Encryption Standard), Triple-
DES, Twofish (a 128-bit block cipher that accepts variable-length key up to 256 bits [105]),
and Rijndael (Advanced Encryption Standard [106]). Table 7.1 shows the results obtained
from processing 1,000,000 blocks. The encryption/decryption speeds (column 2 and 3 in
Table 7.1) are obtained by using the GNU CipherSpeed tool. The CPU usage is computed
based on equation 7.2.
The results demonstrate that secure packet forwarding is quite feasible in wireless
networks as the appropriate cipher only uses about 1.6% of a mobile host’s CPU time.
7.6.2 Overhead of secure roaming support
The computation overhead of the secure roaming support is introduced by the mutual
authentication protocol. The time consumed by different cryptography operations using
the RSA algorithm are shown in Table 7.2. They are obtained by operating 1,000 64-byte
blocks with different keys whose length is 1024 bits. The computation time in one roaming
request can be estimated as follows according to the mutual authentication protocol.
• Mobile host:one signing, one asymmetric decryption, two verifying, and one sym-
metric encryption (whose computation time can be ignored) operations are required.
The computation time is about 90ms.
• Foreign agent:one verifying, one asymmetric encryption, and one signing opera-
tions are required. The computation time is about 50ms.
• Home agent:one verifying and two signing operations are required. The computa-
tion time is about 90ms.
Since roaming is caused by the relative motion between a mobile host and its group
agent, for demonstration purpose, only hosts are moving in the simulations. Figure 7.5
shows the topology of a typical WNMBS. Mobile hosts move in a square area that is
fully covered by 13 base stations. The movement is determined by the random way-point
mobility [56] model. The pause time is 0 second. The maximum speed ranges from 2m/s,
124
Table 7.2Speed of RSA
Operation Signing Verifying Encryption Decryption
Time (ms) 40.73 2.38 2.29 40.66
Base Station
Mobile HostMobile Host Moving AreaBase Station Coverage
GA
Figure 7.5. Topology of a WNMBS
the jogging speed of a person, to 30m/s, the speed of a running vehicle. The radius of
every circle is 250m. Each simulation runs for 5000 seconds. For a mobile host, the mean
interval between two consecutive requests is 416.38 and 56.49 seconds, respectively, when
the maximum speed is 2m/s and 30 m/s.
The rest experiments study the requests related to the group agent GA. Figure 7.6a
shows the frequency of requests as a function of the number of foreign hosts in the area
and their maximum speed, when GA acts as a foreign agent. For 50 foreign hosts, the
125
05
1015
2025
30
50
100
150
200
2500
0.05
0.1
0.15
0.2
Maximum Speed (m/s)
Number of Foreign Mobile Hosts
(a) Number of requests per second as a foreign
agent (do not cache keys)
05
1015
2025
30
50
100
150
200
2500
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
Maximum Speed (m/s)
Number of Foreign Mobile Hosts
(b) Number of requests per second as a foreign
agent (cache keys)
05
1015
2025
30
10
20
30
40
500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Maximum Speed (m/s)
Number of Home Mobile Hosts
(c) Number of requests per second as the home
agent (do not cache keys)
05
1015
2025
30
10
20
30
40
500
0.1
0.2
0.3
0.4
0.5
Maximum Speed (m/s)
Number of Home Mobile Hosts
(d) Number of requests per second as the home
agent (cache keys)
Figure 7.6. Frequency of roaming requests
number of requests per second increases from 0.005 to 0.04 with the maximum speed
increasing from 2m/s to 30m/s. Even with 250 foreign hosts and 30m/s maximum speed,
there are less than 0.2 requests per second. In this set of experiments, the computation
overhead on GA of being a foreign agent is always less than 1% CPU time.
The overhead on GA of being the home agent is determined by the number of hosts
whose home agent is GA and their mobility. Figure 7.6c shows the frequency of requests
as a function of the number of home hosts in the area and the maximum speed. For
50 home hosts and 30m/s maximum speed, the frequency is as high as 0.8 requests per
126
second, because the home agent is involved in every roaming request. In this case, the
computation overhead is about 7.2% CPU time.
The number of requests can be reduced if foreign agents cache the public key of a
mobile host for a period of time. Figure 7.6d shows the results of the experiments in
which foreign agents cache public keys for 200 seconds. The highest frequency is 0.45
requests per second, about a half of that in the previous experiment. The corresponding
computation overhead is about 4% CPU time. The total computation overhead on GA
ranges from 0.2% to 5% CPU time in the experimental study depending on the number of
hosts and their mobility.
7.7 Conclusion
This chapter presents security mechanisms for HMWN to support wireless networks
with movable base stations. The base stations (group agents) serve as a distributed trust
entity. A secure packet forwarding algorithm is designed to protect the network infrastruc-
ture. A protocol is developed to authenticate a mobile host and distribute the group-shared
key. An algorithm is designed to support mobile hosts roaming within the network. To
secure both the foreign group and the mobile host, they mutually authenticate each other
with the help from the home group agent. Experiments have been conducted on a low-end
700MHz PC. The results justify the feasibility of the proposed security mechanisms. The
computation overhead of secure packet forwarding is less than 2% CPU time, and that of
secure roaming support ranges from 0.2% to 5% CPU time depending on the number of
hosts and their motion.
127
8 CONCLUSIONS AND FUTURE WORK
8.1 Conclusions
8.1.1 Study of ad hoc routing protocols
Studying different approaches instead of individual protocols will be of great benefit
to the design and improvement of ad hoc routing protocols. We choose AODV and DSDV
as the representatives of on-demand and proactive approaches. Both protocols utilize dis-
tance vector coupled with destination sequence number, and choose routes in the same
manner. They are differentiated by the way in which they operate (i.e., proactive versus
on-demand). We investigate the performance of DSDV and AODV in terms of packet
delivery ratio, average end-to-end delay, normalized protocol overhead, and normalized
power consumption, under a wide range of network contexts with varied network size,
mobility, and traffic load.
The major observations in this study include:
• Both proactive and on-demand approaches handle topology changes appropriately
as the increase of mobility does not affect much the performance.
• The on-demand approach outperforms the proactive approach in less stressful situa-
tions (i.e., traffic load is light). The proactive approach is more scalable with respect
to traffic load.
• The on-demand approach consumes less power, because it propagates the link break
information faster, thus it avoids sending packets that are dropped eventually.
Although the published results [9, 25] showed that on-demand protocols outperform
proactive protocols and are better suited for mobile ad hoc networks, the proactive proto-
cols provide better support for quality of service (QoS) and anomaly detection. We iden-
128
tify that network congestion is the major reason for performance degradation. Congestion-
aware distance vector (CADV) routing protocol is proposed to address the congestion is-
sue.
In CADV, each routing entry is associated with anexpected delay, which measures
congestion at the next hop. Every host estimates the expected delay based on the mean of
delay for all data packets sent in a past short period of time. When a host broadcasts an
update to neighbors, it specifies the delay it may introduce. A routing decision is made
based on the distance to the destination as well as the expected delay at the next hop.
CADV tries to balance traffic and avoid congestion by giving priority to a route having
low expected delay.
The preliminary study shows CADV outperforms AODV by about 5% in terms of
packet delivery ratio with less protocol overhead.
8.1.2 Study of packet loss in ad hoc networks
Throughput is generally accepted as one of the most important metrics to evaluate
the performance of a routing protocol. It is determined by how many packets have been
sent and how many packets have lost. Studying when and why a packet is dropped will
provide insights in the design of routing and flow control algorithms and the dimensioning
of buffers. We concentrate on congestion-related and mobility-related packet loss.
• Congestion in a network occurs whenever the demands exceed the maximum capac-
ity of a communication link, especially when multiple hosts try to access a shared
media simultaneously.
• Mobility may cause packet loss in different ways. A packet may be dropped at the
source if a route to the destination is not available, or the buffer that stores pending
packets is full. It may also be dropped at an intermediate host if the link to the next
hop has broken.
129
We study the percentages of packet loss due to congestion and mobility in various
network contexts. AODV and DSDV are chosen as representatives of on-demand and
proactive routing protocols respectively. We observe from the experiment results:
• Mobility is the dominant cause for packet loss in AODV, which is responsible for
more than 60% of total packet loss. For DSDV, more than 50% of total packet loss
is congestion-related.
• DSDV loses 10% to 20% more packets than AODV does for UDP traffic. For TCP
traffic, the packet loss for DSDV is a half of that for AODV, because TCP greatly
reduces congestion-related loss.
• Increasing traffic load has a strong impact on packet loss. Mobility decreases packet
loss with light traffic load.
8.1.3 Congestion avoidance routing protocol for ad hoc networks
Congestion control is a problem in ad hoc networks. Compared to the traditional solu-
tions at the transport layer, self-adjusting congestion avoidance (SAGA) routing protocol
is implemented at the network layer. SAGA integrates the channel spatial reuse with the
multi-hop routing to reduce congestion. SAGA is a distance vector routing protocol that
uses intermediate delay (IMD) instead of hop count as the distance. The use of IMD en-
ables routing protocols to select routes that bypass hot spots where contention is intense,
thus enhance the routing performance. It is especially of benefit to networks where topol-
ogy changes are much less frequent than traffic changes. The lazy route query operation
in SAGA uses a special route advertisement for route discovery. Multiple queries can be
included in one advertisement packet to accelerate the establishment of needed routes.
The lazy route query can be applied to other proactive routing protocols that do not have
a dedicated route discovery operation. An approach is provided in SAGA to reduce the
oscillation of the value of IMD and makes the routes stable. SAGA protocol reduces
congestion at every intermediate node. It can be used as a complementary scheme to the
130
end-to-end congestion control/avoidance mechanisms. The intermediate delay obtained
from SAGA can be used to improve the accuracy of round-trip-time (RTT) estimation for
TCP connections.
This research provides methods to estimate the delay at a node using only local infor-
mation. When a node has recent traffic, statistical methods are used to evaluate the mean of
the delay. Otherwise, the underlying MAC protocol is analyzed and probability methods
are applied to compute the expectation of the delay. We analyze the packet transmission
procedure of the distributed coordination function in the IEEE 802.11 standard as a case of
the practical study. These methods are applicable to other contention-based media access
protocols. They can be extended to provide quality of service (QoS) information to upper
layer protocols and applications.
A series of experiments have been conducted to study the performance of routing pro-
tocols under congestion. Two types of UDP traffic as well as the TCP traffic are considered
and the offered traffic load is taken as the input parameter. The maximum moving speed
of nodes and the number of connections are varied. SAGA performs better than DSDV
in all our measurements. A summary of comparison of SAGA with AODV and DSR for
throughput, overhead, and end-to-end delay is as follows.
• SAGA is able to deliver around 90% of the data packets with an offered traffic load
up to 300 kb/s. Its peak throughput is 1.5 to 3.5 times as compared to that of AODV
and DSR.
• Overhead is measured as the ratio of the routing load to the data successfully deliv-
ered to the destination. The overhead of SAGA remains in a range of 15% to 50%.
In similar cases, the overhead of AODV and DSR varies widely and increases fast
as the offered traffic load goes high. The overhead of SAGA is as low as 10% of
that of AODV and 12% of that of DSR in high traffic load.
• For low traffic load, the average end-to-end delay of SAGA is the same as that of
AODV and DSR. When traffic reaches 500 kb/s, the delay of SAGA is 50% less
than that of AODV and 80% less than that of DSR.
131
8.1.4 Wireless networks with movable base stations
The hierarchical mobile wireless network(HMWN) is proposed to support movable
base stations in wireless networks. In a HMWN, mobile hosts are partitioned into groups.
Each group can be viewed as an ad hoc network. It consists of some members and a
group agent that may be a member of another group. The group agent is the representative
of a group. The agent-member relationship forms a hierarchy. A group agent (i.e., a
movable base station) acts as a gateway that connects these two groups. Mobile hosts
belonging to the same group rely on multi-hop routing to communicate with each other.
Communication with a host outside the group is accomplished by the proposed inter-group
routing protocol.
Unlike in the fixed networks, where the location of a host is determined by its network
address, in a mobile network, hosts can move to anywhere without changing the addresses.
In HMWN, the location of a host is the group to which it belongs. The hierarchical mem-
bership management scheme serves two purposes: (a) verifying the identity of a mobile
host for authentication, (b) locating a mobile host for routing protocols. Two kinds of
memberships are maintained by group agents.
• Permanent membership. This is the registration information of a mobile host, such
as public key, billing information, etc. It is established in the bootstrapping phase
and determines if a host can join the system.
• Current membership. This is the location information for a mobile host. The man-
agement requires efficient update schemes to dynamically update it when a mobile
hosts joins, leaves, or roam from one group to another.
As a HMWN is comprised of autonomous groups, the routing protocol must be capable
of accommodating various intra-group routing protocols with least extra overhead. The
approach is to localize the more frequently changing information while disseminating the
less dynamic one. The proposed segmented membership-based group routing (SMGR)
protocol has the following features:
132
• Segmented:Each group routes packets autonomously using its own protocol. When
destination of a packet is outside the group, the packet is sent to the appropriateexit
host.
• Distributed membership-based locating:The exit host of a packet in the group is
identified by querying the membership of the destination.
• Packet encapsulation:The exit host encapsulates the packet to hide the differences
among the routing protocols adopted by different groups. It is also used by security
algorithms.
In SMGR, the topology change (the more dynamic information) is captured and propa-
gated locally within a group by the routing protocol, while the membership change (the
less dynamic information) is distributed to agents following the hierarchy by member-
ship management. Simulation-based experiments demonstrate the scalability of SMGR in
terms of protocol overhead.
8.1.5 Securing wireless networks with movable base stations
In a wireless system, the network infrastructure is protected by ensuring the routing
information will not be forged, modified, or disclosed to an adversary. The packet header
needs to be encrypted as it contains the source, destination, and next hop, which will
expose the relationship among the involved hosts. To reduce encryption/decryption over-
head and the impact of compromised hosts, we propose the security scheme that uses
group shared symmetric key. The scheme consists of two parts:
• A protocol that authenticates a mobile host with its home agent to establish a shared
key.
• An algorithm that cooperates with the routing algorithm to encrypt/decrypt packet
headers. The content of control packets is protected using the same cryptographic
technique. The content of data packets is protected by the upper applications them-
selves.
133
The computation overhead of different cryptographic algorithms is studied through exper-
iments. The result justifies the feasibility of the proposed mechanisms.
In wired environments, Mobile IP is the most widely used protocol to support roaming.
Mobile IP is not an ideal solution for HMWN, because (a) it establishes a “tunnel” between
the home agent and foreign agent, which consumes wireless bandwidth; (2) it does not
support ”group roaming” (i.e., a whole group moves from one place to another). The
essence of roaming support is relocating a mobile host. SMGR protocol naturally supports
roaming as it dynamically locates the destination when forwarding a packet.
Secure roaming support is mandatory for protecting a mobile system. A mutual au-
thentication protocol is developed to authenticate a mobile host with a foreign agent and
establish the shared key. The protocol provides protection to both the foreign group and
the roaming host. The home agent of the host acts as the trust third party in this protocol.
The authentication protocol will fail if the home agent is not available because of
movement or failure. The host will be detached from the system. To make HMWN sur-
vivable from such a single point of failure, the hierarchical fault-tolerant authentication
scheme is applied. A host shares a secret with each of the agents on the path from itself to
the root of the hierarchy. Any of those agents may be the trust third party in the authenti-
cation protocol. The scheme toleratesL − 1 agent failures for a mobile host, whereL is
the height of the hierarchy.
8.2 Future work
The research in this dissertation can be extended in a number of directions. The fol-
lowing summaries some of these directions.
8.2.1 Congestion control in ad hoc networks
Congestion in ad hoc networks can greatly degrade the performance. The set of TCP
congestion control algorithms are based on the principle of conservation of packets. In ad
hoc networks, the existence of multiple routes between two nodes provides an opportunity
134
for the routing protocols to select appropriate ones. A cross-layer design integrating the
MAC, network, and transport layers will provide solutions to the congestion control prob-
lem. Based upon the research on the congestion avoidance routing protocol in this disser-
tation, the following research questions need investigation: How routing protocols can use
the directional transmission provide by smart antennae to reduce channel interference and
contention? How to utilize the multi-rate, multi-range, and multi-channel supports from
IEEE 802.11 standard to minimize congestion? What is the tradeoff between connectivity
and congestion avoidance? What are the advantages and disadvantages of different error-
detection strategies, e.g. network detection and end-node detection, when they are used
to infer congestion? This research will contribute to the development of adaptive protocol
suites for ad hoc networks. It will be of benefit in advancing perversive computing and
communication.
8.2.2 Trusted communication
Secure and trusted collaboration over worldwide computer networks will enable the
formation of trusted global partnership in education, research, and with applications to
business, military, security of the nation. Trusted communication is a necessity for trusted
collaboration. Although security mechanisms can be applied to protect the communica-
tion between two participants, the collaboration is vulnerable to untrustworthy behaviors.
In the collaborative network, every collaborator participates with others to deliver infor-
mation. The safety of a communication solely depends on a proper choice of a sequence
of collaborators to reach the destination. Sending information through a path that only
involves trustworthy participants will decrease the probability of malicious attacks and
information leakage. To investigate the problem of using trust to estimate the risk of se-
lecting a collaborator, we need to (a) formalize trust for communication by building a
model that quantifies the trustworthiness of a collaborator based on its behaviors, relia-
bility, and security; (b) develop algorithms to assess the trustworthiness of a path based
on information of collaborators; (c) design protocols that propagate trust information and
135
discover paths according to specific requirements; (d) experimentally study the integra-
tion of security mechanisms such as authentication, encryption/decryption, and filtering
to defend against malicious attacks.
8.2.3 Privacy-preserved communication
The increasing amount of data sharing and collaboration calls for privacy-preserving
mechanisms. Existing research efforts have studied the anonymous communication prob-
lem by hiding the identity of the subject in a group of participants. The proposed schemes
ensure that the source of a communication is unknown, but everybody may know the con-
tent. Another approach for the privacy preservation problem is to remove the association
between the content of the communication and the identity of the source. Somebody may
know the source while others may know the content, but nobody knows both. The ap-
proaches will use trusted proxies to protect privacy in a dynamic communication environ-
ment. Research questions include: How to establish and maintain the trust relationships?
How to measure the level of privacy that a specific approach can achieve? What are the
tradeoffs for achieving a certain level privacy? What are the possible attacks and security
threats?
LIST OF REFERENCES
136
LIST OF REFERENCES
[1] C. E. Perkins and P. Bhagwat, “Highly dynamic destination-sequenced distance-vector routing (DSDV) for mobile computers,” inProceedings of the ACM Confer-ence on Communications Architectures, Protocols and Applications (SIGCOMM),(London, United Kingdom), pp. 234–244, August 1994.
[2] C. E. Perkins, E. M. Belding-Royer, and S. R. Das, “Ad hoc on-demand distancevector (AODV) routing,”RFC 3561, http://www.ietf.org/rfc/rfc3561.txt, July 2003.
[3] D. B. Johnson and D. A. Maltz, “Dynamic source routing in ad hoc wireless net-works,” inMobile Computing(T. Imielinski and H. Korth, eds.), ch. 5, pp. 153–181,Kluwer Academic Publishers, 1996.
[4] H.-Y. Hsieh and R. Sivakumar, “On using the ad-hoc network model in cellularpacket data networks,” inProceedings of the 3rd ACM International Symposiumon Mobile Ad Hoc Networking & Computing (MobiHoc), (Lausanne, Switzerland),pp. 36–47, June 2002.
[5] H. Wu, C. Qiao, S. De, and O. Tonguz, “Integrated cellular and ad hoc relayingsystem: iCAR,”IEEE Journal on Selected Area in Communications, vol. 19, no. 10,pp. 2105–2115, 2001.
[6] X. Wu, B. Mukherjee, and G.-H. Chan, “MACA: An efficient channel allocationscheme in cellular network,” inProceedings of GlobeCom 2000, vol. 3, (San Fran-cisco), pp. 1385–1389, November 2000.
[7] “WaveLAN/PCMCIA card user’s guide, Lucent Technologies,” October 1996.
[8] C. E. Perkins, E. M. Royer, S. R. Das, and M. K. Marine, “Performance comparisonof two on-demand routing protocols for ad hoc networks,”IEEE Personal Commu-nications, vol. 8, no. 1, pp. 16–28, 2001.
[9] J. Broch, D. A. Maltz, D. B. Johnson, Y. Hu, and J. Jetcheva, “A performance com-parison of multi-hop wireless ad hoc network routing protocols,” inProceedings ofthe 4th ACM/IEEE International Conference on Mobile Computing and Networking(MobiCom), (Dallas), pp. 85–97, October 1998.
[10] Y. Lu, W. Wang, Y. Zhong, and B. Bhargava, “Performance study of two distancevector routing protocols for mobile ad hoc networks,” Tech. Rep. CSD 02-016,Department of Computer Sciences, Purdue University, 2002.
[11] J. J. Garcia-Luna-Aceves and M. Spohn, “Source-tree routing in wireless net-works,” in Proceedings of 7th International Conference on Network Protocols(ICNP), pp. 273–282, November 1999.
137
[12] J. J. Garcia-Luna-Aceves and M. Spohn, “Transmission-efficient routing in wirelessnetworks using link-state information,”Mobile Networks and Applications, vol. 6,no. 3, pp. 223–238, 2001.
[13] M. R. Pearlman, Z. J. Haas, and S. I. Mir, “Using routing zones to support routemaintenance in ad hoc networks,” inProceedings of Wireless Communications andNetworking Conference (WCNC), September 2000.
[14] Z. J. Haas, M. R. Pearlman, and P. Samar, “The zone routing protocol (ZRP) for adhoc networks,”IETF Internet Draft. http://www.ietf.org/internet-drafts/draft-ietf-manet-zone-zrp-02.txt, July 2002.
[15] R. Sivakumar, P. Sinha, and V. Bharghavan, “CEDAR: A core-extraction distributedad hoc routing algorithm,”IEEE Journal on Selected Areas in Communications,vol. 17, pp. 1454–1465, 1999.
[16] S. Corson and J. Macker, “Mobile ad hoc networking (MANET): Rout-ing protocol performance issues and evaluation considerations,”RFC 2501,http://www.ietf.org/rfc/rfc2501.txt, January 1999.
[17] S. R. Das, R. C. neda, and J. Yan, “Simulation-based performance evaluation ofrouting protocols for mobile ad hoc networks,”Mobile Networks and Applications,vol. 5, pp. 179–189, 2000.
[18] “The network simulator – ns-2,”http://www.isi.edu/nsnam/ns/.
[19] H. Jiang and J. J. Garcia-Luna-Aceves, “Performance comparison of three routingprotocols for ad hoc networks,” inProceedings of IEEE ICCCN 2001, (Phoenix),pp. 547–554, October 2001.
[20] X. Zeng, R. Bagrodia, and M. Geria, “GloMoSim: A library for parallel simulationof large-scale wireless networks,” inProceedings of the 12th Workshop on Paralleland Distributed Simulations (PADS), (Banff, Alberta, Canada), May 1998.
[21] T. Camp, J. Boleng, B. Williams, L. Wilcox, and W. Navidi, “Performance compar-ison of two location based routing protocols for ad hoc networks,” inProceedingsof IEEE INFOCOM, vol. 3, (New York), pp. 1678–1687, June 2002.
[22] R. V. Boppana and S. P. Konduru, “An adaptive distance vector routing algorithmfor mobile, ad hoc networks,” inProceedings of IEEE INFOCOM, vol. 3, (Anchor-age, Alaska), pp. 1753–1762, April 2001.
[23] N. R. Draper and H. Smith,Applied Regression Analysis (Third Edition). JohnWiley & Sons, Inc., 1998.
[24] R. Kravets and P. Krishnan, “Application-driven power management for mobilecommunication,”Wireless Networks, vol. 6, no. 4, pp. 263–277, 2000.
[25] P. Johansson, T. Larsson, N. Hedman, B. Mielczarek, and M. Degermark,“Scenario-based performance analysis of routing protocols for mobile ad-hoc net-works,” in Proceedings of the 5th ACM/IEEE International Conference on MobileComputing and Networking (MobiCom), (Seattle), pp. 195–206, August 1999.
138
[26] W. Wang, Y. Lu, and B. Bhargava, “On vulnerability and protection of ad hoc on-demand distance vector protocol,” inProceedings of International Conference onTelecommunications (ICT), (Papeete, French Polynesia), pp. 375–382, February2003.
[27] W. Wang, Y. Lu, and B. Bhargava, “On security study of two distance vector routingprotocols for mobile ad hoc networks,” inProceedings of IEEE International Con-ference on Pervasive Computing and Communications (PerCom), (Texas), pp. 179–186, March 2003.
[28] I. Cidon, A. Khamisy, and M. Sidi, “Analysis of packet loss process in high-speednetworks,”IEEE Transactions on Information Theory, vol. 39, pp. 98–108, January1993.
[29] J.-C. Bolot, “End-to-end packet delay and loss behavior in the internet,” inProceed-ings of Conference on Communications Architectures, Protocols and Applications(SIGCOMM), (San Francisco), pp. 289–298, September 1993.
[30] S. Biaz and N. H. Vaidya, “Distinguishing congestion losses from wireless trans-mission losses: A negative result,” inProceedings of IEEE ICCCN, (New Orleans),pp. 722–731, October 1998.
[31] F. Anjum and L. Tassiulas, “On the behavior of different TCP algorithms overa wireless channel with correlated packet losses,” inProceedings of the Interna-tional Conference on Measurement and Modeling of Computer Systems, (Atlanta),pp. 155–165, 1999.
[32] T. V. Lakshman, U. Madhow, and B. Suter, “TCP/IP performance with randomloss and bidirectional congestion,”IEEE/ACM Transactions on Networking (TON),vol. 8, no. 5, pp. 541–555, 2000.
[33] M. K. Marina and S. R. Das, “On-demand multipath distance vector routing in adhoc networks,” inProceedings of the International Conference on Network Proco-tols (ICNP), (Riverside), pp. 14–23, November 2001.
[34] T. Goff, N. B. Abu-Ghazaleh, D. S. Phatak, and R. Kahvecioglu, “Preemptive rout-ing in ad hoc networks,” inProceedings of the 7th International Conference onMobile Computing and Networking (MobiCom), July 2001.
[35] M. Grossglauser and D. Tse, “Mobility increases the capacity of ad-hoc wirelessnetworks,” in Proceedings of IEEE INFOCOM, vol. 3, (Anchorage), pp. 1360–1369, April 2001.
[36] V. Jacobson, “Congestion avoidance and control,” inProceedings of Symposium onCommunications Architectures and Protocols (SIGCOMM), (Stanford), pp. 314–329, August 1988.
[37] L. S. Brakmo, S. W. O’Malley, and L. L. Peterson, “TCP Vegas: New techniques forcongestion detection and avoidance,” inProceedings of the Conference on Commu-nications Architectures, Protocols and Applications (SIGCOMM), (London, UnitedKingdom), pp. 24–35, August 1994.
[38] J. Mo, R. La, V. Anantharam, and J. Walrand, “Analysis and comparison of TCPReno and Vegas,” inProceedings of INFOCOM, pp. 1556–1563, March 1999.
139
[39] J. Padhye, V. Firoiu, D. F. Towsley, and J. F. Kurose, “Modeling TCP Reno perfor-mance: A simple model and its empirical validation,”IEEE/ACM Transactions onNetworking (TON), vol. 8, no. 2, pp. 133–145, 2000.
[40] J. Liu and S. Singh, “ATCP: TCP for mobile ad hoc networks,”IEEE Journal onSelected Areas in Communications, vol. 19, no. 7, pp. 1300–1315, 2001.
[41] C. Parsa and J. J. Garcia-Luna-Aceves, “Improving TCP performance over wirelessnetworks at the link layer,”Mobile Networks and Applications, vol. 5, no. 1, pp. 57–71, 2000.
[42] K. Chandran, S. Raghunathan, S. Venkatesan, and R. Prakash, “A feedback-basedscheme for improving TCP performance in ad hoc wireless networks,”IEEE Per-sonal Communications, vol. 8, no. 1, pp. 34–39, 2001.
[43] C. Zhang and V. Tsaoussidis, “TCP-Real: Improving real-time capabilities of TCPover heterogeneous networks,” inProceedings of the 11th International Workshopon Network and Operating Systems Support for Digital Audio and Video (NOSS-DAV), (Port Jefferson), pp. 189–198, June 2001.
[44] F. Wang and Y. Zhang, “Improving TCP performance over mobile ad-hoc networkswith out-of-order detection and response,” inProceedings of the 3rd ACM Interna-tional Symposium on Mobile Ad Hoc Networking & Computing (MobiHoc), (Lau-sanne, Switzerland), pp. 217–225, June 2002.
[45] S. Cen, P. C. Cosman, and G. M. Voelker, “End-to-end differentiation of congestionand wireless losses,”IEEE/ACM Transactions on Networking (TON), vol. 11, no. 5,pp. 703–717, 2003.
[46] S. Kunniyur and R. Srikant, “End-to-end congestion control schemes: Utility func-tions, random losses and ECN marks,”IEEE/ACM Transactions on Networking(TON), vol. 11, no. 5, pp. 689–702, 2003.
[47] C.-K. Toh, “Associativity-based routing for ad-hoc mobile networks,”Wireless Per-sonal Communications Journal, vol. 4, no. 2, pp. 103–139, 1997.
[48] H. Hassanein and A. Zhou, “Routing with load balancing in wireless ad hoc net-works,” inProceedings of the 4th ACM International Workshop on Modeling, Anal-ysis and Simulation of Wireless and Mobile Systems, (Rome, Italy), pp. 89–96, July2001.
[49] S.-J. Lee and M. Gerla, “Dynamic load-aware routing in ad hoc networks,” inPro-ceedings of ICC, vol. 10, pp. 3206–3210, June 2001.
[50] A. A. Bertossi and M. A. Bonuccelli, “Code assignment for hidden terminal inter-ference avoidance in multihop packet radio networks,”IEEE/ACM Transactions onNetworking (TON), vol. 3, no. 4, pp. 441–449, 1995.
[51] Y. Lu, Y. Zhong, and B. Bhargava, “Packet loss in mobile ad hoc networks,” Tech-nical Report CSD 03-009, Department of Computer Sciences, Purdue University,2003.
[52] “ANSI/IEEE Std 802.11, 1999 Edition,”http://standards.ieee.org/getieee802/download/802.11-1999.pdf.
140
[53] H. Luo, P. Medvedev, and S. Lu, “A self-coordinating approach to distributed fairqueueing in ad hoc wireless networks,” inProceedings of IEEE INFOCOM, vol. 3,(Anchorage), pp. 1370–1379, April 2001.
[54] E. Gelenbe and G. Pujolle,Introduction to Queueing Networks, Second Edition.John Wiley & Sons, 1998.
[55] W. Mendenhall and R. J. Beaver,Introduction to Probability and Statistics. Boston:PWS-Kent Pub. Co., 8th ed., 1991.
[56] Y. Lu, W. Wang, Y. Zhong, and B. Bhargava, “Study of distance vector routingprotocols for mobile ad hoc networks,” inProceedings of IEEE International Con-ference on Pervasive Computing and Communications (PerCom), (Texas), pp. 187–194, March 2003.
[57] J. Jaffe and F. Moss, “A responsive distributed routing algorithm for computer net-works,” IEEE Transactions on Communications, vol. 30, no. 7, pp. 1758–1762,1982.
[58] D. S. J. De Couto, D. Aguayo, J. Bicket, and R. Morris, “A high-throughput pathmetric for multi-hop wireless routing,” inProceedings of the 9th International Con-ference on Mobile Computing and Networking (MobiCom), (San Diego), pp. 134–146, September 2003.
[59] Y. Lu and B. Bhargava, “Self-adjusting congestion avoidance routing protocol forad hoc networks,” Technical Report CSD 03-018, Department of Computer Sci-ences, Purdue University, 2003.
[60] W. Willinger, M. S. Taqqu, R. Sherman, and D. V. Wilson, “Self-similarity throughhigh-variability: Statistical analysis of ethernet LAN traffic at the source level,”IEEE/ACM Transactions on Networking (TON), vol. 5, no. 1, pp. 71–86, 1997.
[61] A. Tsirigos and Z. Haas, “Multipath routing in the presence of frequent topologicalchanges,”IEEE Communications Magazine, vol. 39, no. 11, pp. 132–138, 2001.
[62] S. K. Das, A. Mukherjee, S. Bandyopadhyay, D. Saha, and K. Paul, “An adaptiveframework for QoS routing through multiple paths in ad hoc wireless networks,”Journal of Parallel and Distributed Computing, vol. 63, no. 2, pp. 141–153, 2003.
[63] A. Boukerche, S. K. Das, and A. Fabbri, “Message traffic control capabilities ofthe R-DSDV protocol in mobile ad hoc networks,” inProceedings of the 4th ACMInternational Workshop on Modeling, Analysis and Simulation of Wireless and Mo-bile Systems, (Rome, Italy), pp. 105–112, 2001.
[64] A. Boukerche and S. K. Das, “Congestion control performance of R-DSDV pro-tocol in multihop wireless ad hoc networks,”Wireless Networks, vol. 9, no. 3,pp. 261–270, 2003.
[65] D. S. J. De Couto, D. Aguayo, B. A. Chambers, and R. Morris, “Performance ofmultihop wireless networks: Shortest path is not enough,”ACM SIGCOMM Com-puter Communication Review, vol. 33, no. 1, pp. 83–88, 2003.
[66] C. Cordeiro, S. R. Das, and D. P. Agrawal, “COPAS: Dynamic contention-balancing to enhance the performance of TCP over multi-hop wireless networks,”in Proceedings of the 10th International Conference on Computer Communicationand Networks (IC3N), (Miami), pp. 382–387, October 2002.
141
[67] S. Xu and T. Saadawi, “Revealing the problems with 802.11 medium access controlprotocol in multi-hop wireless ad hoc networks,”Journal of Computer Networks,vol. 38, no. 4, pp. 531–548, 2002.
[68] S.-T. Sheu, Y. Tsai, and J. Chen, “MR2RP: The multi-rate and multi-range routingprotocol for IEEE 802.11 ad hoc wireless networks,”Wireless Networks, vol. 9,no. 2, pp. 165–177, 2003.
[69] S. Chen and K. Nahrstedt, “Distributed quality-of-service routing in ad hoc net-works,” IEEE Journal on Selected Areas in Communications, vol. 17, no. 8,pp. 1488–1505, 1999.
[70] C. Lin, “On-demand QoS routing in multihop mobile networks,” inProceedings ofIEEE INFOCOM, vol. 3, (Anchorage), pp. 1735–1744, April 2001.
[71] C. Zhu and M. Corson, “QoS routing for mobile ad hoc networks,” inProceedingsof IEEE INFOCOM, vol. 2, (New York), pp. 958–967, June 2002.
[72] Y. Zhang and W. Li, “An integrated environment for testing mobile ad-hoc net-works,” in Proceedings of the 3rd ACM International Symposium on Mobile AdHoc Networking & Computing (MobiHoc), (Lausanne, Switzerland), pp. 104–111,June 2002.
[73] Y.-C. Tseng, Y.-F. Li, and Y.-C. Chang, “On route lifetime in multihop mobile adhoc networks,”IEEE Transactions on Mobile Computing, vol. 2, no. 4, pp. 366–376, 2003.
[74] “IEEE Std 802.11a-1999 (supplement to IEEE Std 802.11-1999),”http://standards.ieee.org/getieee802/download/802.11a-1999.pdf.
[75] C. Santivanez, B. McDonald, I. Stavrakakis, and R. Ramanathan, “On the scalabil-ity of ad hoc routing protocols,” inProceedings of IEEE INFOCOM, vol. 3, (NewYork), pp. 1688–1697, June 2002.
[76] X. Hong, M. Gerla, G. Pei, and C. Chiang, “A group mobility model for adhoc wireless networks,” inProceedings of the 2nd ACM International Workshopon Modeling, Analysis and Simulation of Wireless and Mobile Systems, (Seattle),pp. 53–60, August 1999.
[77] K. B. Egevang, C. Communications, and P. Francis, “The IP network address trans-lator (nat),”RFC 1631, http://www.ietf.org/rfc/rfc1631.txt, May 1994.
[78] Y.-D. Lin and Y.-C. Hsu, “Multihop cellular: A new architecture for wirelesscommunications,” inProceedings of IEEE INFOCOM, vol. 3, (Tel-Aviv, Israel),pp. 1273–1282, March 2000.
[79] H. Luo, R. Ramjee, P. Sinha, L. E. Li, and S. Lu, “UCAN: a unified cellular andad-hoc network architecture,” inProceedings of the 9th ACM International Confer-ence on Mobile Computing and Networking (MobiCom), (San Diego), pp. 353–367,September 2003.
[80] S. Nesargi and R. Prakash, “Distributed wireless channel allocation in networkswith mobile base stations,”IEEE Transactions on Vehicular Technology, vol. 51,no. 6, pp. 1407–1421, 2002.
142
[81] R. Ramanathan and S. Steenstrup, “Hierarchically-organized, multihop mobilewireless networks for quality-of-service support,”Mobile Networks and Applica-tions, vol. 3, no. 1, pp. 101–119, 1998.
[82] S. Banerjee and S. Khuller, “A clustering scheme for hierarchical control in multi-hop wireless networks,” inProceedings of IEEE INFOCOM, vol. 2, (Anchorage),pp. 1028–1037, April 2001.
[83] B. Liang and Z. J. Haas, “Virtual backbone generation and maintenance in ad hocnetwork mobility management,” inProceedings of IEEE INFOCOM, (Tel Aviv,Israel), March 2000.
[84] Y. Chang and C. Hsu, “Routing in wireless/mobile ad-hoc networks via dynamicgroup construction,”Mobile Networks and Applications, vol. 5, no. 1, pp. 27–37,2000.
[85] I. Rubin, A. Behzad, R. Zhang, H. Luo, and E. Caballero, “TBONE: A mobile-backbone protocol for ad hoc wireless networks,” inProceedings of IEEEAerospace Conference, vol. 6, (Big Sky, MT), pp. 2727–2740, March 2002.
[86] “IETF MANET working group,”http://www.ietf.org/html.charters/manet-charter.html.
[87] N. Borisov, I. Goldberg, and D. Wagner, “Intercepting mobile communications:The insecurity of 802.11,” inProceedings of the 7th ACM International Conferenceon Mobile Computing and Networking (MobiCom), (Rome, Italy), pp. 180–189,July 2001.
[88] V. Bharghavan, “Secure wireless LANs,” inProceedings of the 2nd ACM Con-ference on Computer and Communications Security (CCS), (Fairfax), pp. 10–17,1994.
[89] S. H. Park, A. Ganz, and Z. Ganz, “Security protocol for IEEE 802.11 wirelesslocal area network,”Mobile Networks and Applications, vol. 3, no. 3, pp. 237–246,1998.
[90] L. Zhou and Z. J. Haas, “Securing ad hoc networks,”IEEE Network Magazine,vol. 13, no. 6, 1999.
[91] J. Binkley and W. Trost, “Authenticated ad hoc routing at the link layer for mobilesystems,”Wireless Networks, vol. 7, no. 2, pp. 139–145, 2001.
[92] B. Awerbuch, D. Holmer, C. Nita-Rotaru, and H. Rubens, “An on-demand securerouting protocol resilient to byzantine failures,” inProceedings of Workshop onWireless Security (WiSe), (Atlanta), September 2002.
[93] M. G. Zapata and N. Asokan, “Securing ad hoc routing protocols,” inProceedingsof Workshop on Wireless Security (WiSe), (Atlanta), September 2002.
[94] J. Kong, P. Zerfos, H. Luo, S. Lu, and L. Zhang, “Providing robust and ubiquitoussecurity support for mobile ad-hoc networks,” inProceedings of 9th InternationalConference on Network Protocols (ICNP), (Riverside), pp. 251–260, November2001.
143
[95] Y.-C. Hu, A. Perrig, and D. B. Johnson, “Ariadne: A secure on-demand routingprotocol for ad hoc networks,” inProceedings of the 8th International Conferenceon Mobile Computing and Networking (MobiCom), (Atlanta), pp. 12–23, 2002.
[96] B. Schneier,Applied Cryptography Second Edition : Protocols, Algorithms, andSource Code in C. John Wiley & Sons, Inc., 1996.
[97] J. Chuprun, C. Bergstrom, and A. Guzek, “Advanced interference rejection andanti-jam methods for low power mobile battlefield communications,” inProceed-ings of IEEE Military Communications Conference (MILCOM 97), vol. 2, pp. 841–846, 1997.
[98] W. Myrick, J. Goldstein, and M. Zoltowski, “Low complexity anti-jam space-timeprocessing for gps,” inProceedings of IEEE International Conference on Acoustics,Speech, and Signal Processing (ICASSP ’01), vol. 4, (Salt Lake City), pp. 2233–2236, May 2001.
[99] F. Buchholz, T. E. Daniels, B. Kuperman, and C. Shields, “Packet tracker finalreport,” Tech. Rep. TR 2000-23, CERIAS, Purdue University, 2000.
[100] M. Burrows, M. Abadi, and R. Needham, “A logic of authentication,”ACM Trans-actions on Computer Systems, vol. 8, no. 1, 1990.
[101] Y. Lu, B. Bhargava, W. Wang, Y. Zhong, and X. Wu, “Secure wireless networkwith movable base stations,”IEICE Transactions on Communications, IEICE/IEEEJoint Special Issue on Assurance Systems and Networks, vol. E86-B, pp. 2922–2930, October 2003.
[102] B. Bhargava, S. Kamisetty, and S. Madria, “Fault-tolerant authentication in mobilecomputing,” in Proceedings of International Conference on Internet Computing(IC), (Las Vegas), June 2000.
[103] D. McClure and B. Bhargava, “On assigning priorities of keying parameters in asecure mobile network,” inProceedings of IEEE Workshop on “Reliable and SecureApplication in Mobile Environment”, (New Orleans), October 2001.
[104] “Packet length distributions,”http://www.caida.org/analysis/AIX/plenhist/.
[105] B. Schneier, J. Kelsey, D. Whiting, D. Wagner, and C. Hall, “Twofish: A 128-bitblock cipher,”http://www.counterpane.com/twofish.pdf, June 1998.
[106] J. Daemen and V. Rijmen,The Design of Rijndael. Information Security and Cryp-tography SE, Springer-Verlag New York, Inc., 2002.
VITA
144
VITA
Yi Lu received the Doctor of Philosophy degree in computer science from Purdue
University, West Lafayette in August 2004. He got the master’s degree from Institute of
Software, Chinese Academy of Sciences in 1999, and the bachelor’s degree from Univer-
sity of Science and Technology of China in 1996, both in computer science. His research
interests include wireless network security, heterogeneous wireless networks, routing and
congestion control protocols for ad hoc networks, and trust modeling for peer-to-peer ap-
plications. He is a member of IEEE and ACM, and a member of IEEE Computer Society.